[Inductor][Grouped Gemm] Add Blackwell CuTeDSL Kernel#167340
Closed
NikhilAPatel wants to merge 1 commit intopytorch:mainfrom
Closed
[Inductor][Grouped Gemm] Add Blackwell CuTeDSL Kernel#167340NikhilAPatel wants to merge 1 commit intopytorch:mainfrom
NikhilAPatel wants to merge 1 commit intopytorch:mainfrom
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/167340
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 44e3539 with merge base b91a2ab ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@NikhilAPatel has exported this pull request. If you are a Meta employee, you can view the originating Diff in D86537373. |
jananisriram
approved these changes
Nov 7, 2025
NikhilAPatel
force-pushed
the
export-D86537373
branch
from
6e62dae to
4b39d15
Compare
NikhilAPatel
added a commit
to NikhilAPatel/pytorch
that referenced
this pull request
Nov 7, 2025
Summary: This is a reland of D82010227, which previously contained a minor bug in the logic that determined whether the kernel should be enabled. As a result, it was incorrectly activated on non-Blackwell GPUs. Test Plan: Inductor test (fbcode): `INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1"` Tritonbench (fbcode): `clear; CUDA_VISIBLE_DEVICES=7 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 buck2 run mode/opt //pytorch/tritonbench:run -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1" -- --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_cute_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Tritonbench(oss): `clear; CUDA_VISIBLE_DEVICES=2 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 python run.py --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_triton_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Unit Tests(oss): `clear; python test/inductor/test_cutedsl_grouped_mm.py` Reviewed By: jananisriram Differential Revision: D86537373
NikhilAPatel
force-pushed
the
export-D86537373
branch
from
4b39d15 to
408a499
Compare
NikhilAPatel
added a commit
to NikhilAPatel/pytorch
that referenced
this pull request
Nov 8, 2025
Summary: This is a reland of D82010227, which previously contained a minor bug in the logic that determined whether the kernel should be enabled. As a result, it was incorrectly activated on non-Blackwell GPUs. Test Plan: Inductor test (fbcode): `INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1"` Tritonbench (fbcode): `clear; CUDA_VISIBLE_DEVICES=7 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 buck2 run mode/opt //pytorch/tritonbench:run -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1" -- --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_cute_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Tritonbench(oss): `clear; CUDA_VISIBLE_DEVICES=2 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 python run.py --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_triton_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Unit Tests(oss): `clear; python test/inductor/test_cutedsl_grouped_mm.py` Reviewed By: jananisriram Differential Revision: D86537373
Summary: This is a reland of D82010227, which previously contained a minor bug in the logic that determined whether the kernel should be enabled. As a result, it was incorrectly activated on non-Blackwell GPUs. Test Plan: Inductor test (fbcode): `INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1"` Tritonbench (fbcode): `clear; CUDA_VISIBLE_DEVICES=7 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 buck2 run mode/opt //pytorch/tritonbench:run -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1" -- --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_cute_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Tritonbench(oss): `clear; CUDA_VISIBLE_DEVICES=2 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 python run.py --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_triton_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Unit Tests(oss): `clear; python test/inductor/test_cutedsl_grouped_mm.py` Reviewed By: jananisriram Differential Revision: D86537373
NikhilAPatel
force-pushed
the
export-D86537373
branch
from
408a499 to
44e3539
Compare
Contributor
|
@pytorchbot merge (Initiating merge automatically since Phabricator Diff has merged) |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
ashahab
pushed a commit
to ashahab/pytorch
that referenced
this pull request
Nov 14, 2025
Summary: This is a reland of pytorch#165036, which previously contained a minor bug in the logic that determined whether the kernel should be enabled. As a result, it was incorrectly activated on non-Blackwell GPUs. Test Plan: Inductor test (fbcode): `INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1"` Tritonbench (fbcode): `clear; CUDA_VISIBLE_DEVICES=7 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 buck2 run mode/opt //pytorch/tritonbench:run -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1" -- --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_cute_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Tritonbench(oss): `clear; CUDA_VISIBLE_DEVICES=2 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 python run.py --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_triton_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Unit Tests(oss): `clear; python test/inductor/test_cutedsl_grouped_mm.py` Differential Revision: D86537373 Pull Request resolved: pytorch#167340 Approved by: https://github.com/jananisriram
ashahab
pushed a commit
to ashahab/pytorch
that referenced
this pull request
Nov 14, 2025
Summary: This is a reland of pytorch#165036, which previously contained a minor bug in the logic that determined whether the kernel should be enabled. As a result, it was incorrectly activated on non-Blackwell GPUs. Test Plan: Inductor test (fbcode): `INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1"` Tritonbench (fbcode): `clear; CUDA_VISIBLE_DEVICES=7 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 buck2 run mode/opt //pytorch/tritonbench:run -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1" -- --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_cute_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Tritonbench(oss): `clear; CUDA_VISIBLE_DEVICES=2 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 python run.py --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_triton_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Unit Tests(oss): `clear; python test/inductor/test_cutedsl_grouped_mm.py` Differential Revision: D86537373 Pull Request resolved: pytorch#167340 Approved by: https://github.com/jananisriram
ashahab
pushed a commit
to ashahab/pytorch
that referenced
this pull request
Nov 15, 2025
Summary: This is a reland of pytorch#165036, which previously contained a minor bug in the logic that determined whether the kernel should be enabled. As a result, it was incorrectly activated on non-Blackwell GPUs. Test Plan: Inductor test (fbcode): `INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1"` Tritonbench (fbcode): `clear; CUDA_VISIBLE_DEVICES=7 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 buck2 run mode/opt //pytorch/tritonbench:run -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1" -- --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_cute_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Tritonbench(oss): `clear; CUDA_VISIBLE_DEVICES=2 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 python run.py --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_triton_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Unit Tests(oss): `clear; python test/inductor/test_cutedsl_grouped_mm.py` Differential Revision: D86537373 Pull Request resolved: pytorch#167340 Approved by: https://github.com/jananisriram
Silv3S
pushed a commit
to Silv3S/pytorch
that referenced
this pull request
Nov 18, 2025
Summary: This is a reland of pytorch#165036, which previously contained a minor bug in the logic that determined whether the kernel should be enabled. As a result, it was incorrectly activated on non-Blackwell GPUs. Test Plan: Inductor test (fbcode): `INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1"` Tritonbench (fbcode): `clear; CUDA_VISIBLE_DEVICES=7 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 buck2 run mode/opt //pytorch/tritonbench:run -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1" -- --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_cute_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Tritonbench(oss): `clear; CUDA_VISIBLE_DEVICES=2 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 python run.py --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_triton_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracy` Unit Tests(oss): `clear; python test/inductor/test_cutedsl_grouped_mm.py` Differential Revision: D86537373 Pull Request resolved: pytorch#167340 Approved by: https://github.com/jananisriram
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary: This is a reland of #165036, which previously contained a minor bug in the logic that determined whether the kernel should be enabled. As a result, it was incorrectly activated on non-Blackwell GPUs.
Test Plan:
Inductor test (fbcode):
INDUCTOR_TEST_DISABLE_FRESH_CACHE=1 TORCHINDUCTOR_CACHE_DIR=~/cutetest buck2 run mode/opt //caffe2/test/inductor:cutedsl_grouped_mm -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1"Tritonbench (fbcode):
clear; CUDA_VISIBLE_DEVICES=7 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 buck2 run mode/opt //pytorch/tritonbench:run -c fbcode.nvcc_arch=b200a -c fbcode.enable_gpu_sections=true -c fbcode.platform010_cuda_version=12.8 -m "ovr_config//third-party/pypi/nvidia-cutlass-dsl/constraints:4.2.1" -- --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_cute_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracyTritonbench(oss):
clear; CUDA_VISIBLE_DEVICES=2 TRITON_PRINT_AUTOTUNING=1 TRITON_ALWAYS_COMPILE=1 TORCH_LOGS=+inductor TORCHINDUCTOR_FORCE_DISABLE_CACHES=1 TORCHINDUCTOR_MAX_AUTOTUNE_GEMM=1 python run.py --op grouped_gemm --only aten_grouped_mm,preprocessed_pt2_triton_grouped_mm --precision bf16 --num-inputs 1 --metrics tflops,accuracyUnit Tests(oss):
clear; python test/inductor/test_cutedsl_grouped_mm.pyDifferential Revision: D86537373
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben