Contiguous subgraph decomposition by exclamaforte · Pull Request #161241 · pytorch/pytorch

exclamaforte · 2025-08-22T03:25:19Z

Summary

Adds a subgraph decomposition for addmm and mm that performs well on large K compared to M and N, and functions well as an alternative to split-k on AMD (transposed only), which does not support AMD currently.

Background

On AMD (MI300x), for a matmul A * B, if B is non-contiguous, the resulting matmul is quite a bit slower.
For example:

  args[0]: TensorBox(StorageBox(
    InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1]))
  ))
  args[1]: TensorBox(StorageBox(
    InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[1, 178176]))
  ))

is a lot slower than:

  args[0]: TensorBox(StorageBox(
    InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1]))
  ))
  args[1]: TensorBox(StorageBox(
    InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[6144, 1]))
  ))

This PR adds a subgraph decomposition to test out whether making B contiguous is faster than just using the normal kernels.

Data

I ran this on unique non-contiguous shapes from torchbench/huggingface and got these speedups:

Parsed 420 unique shapes from benchmark output
addmm improvements when best:
  addmm_16448x512x2048: +0.14%
  addmm_128x2048x2048: +0.01%
  addmm_128x768x1000: +0.75%
  addmm_12672x3072x768: +1.08%
  addmm_512x768x32000: +0.62%
  addmm_12608x384x384: +0.00%
  addmm_4160x1024x4096: +0.90%
  addmm_16x768x2: +0.56%
  addmm_12608x3072x768: +0.09%
  addmm_64x4096x1000: +2.77%
  addmm_256x1024x512: +1.99%
  addmm_30x256x256: +1.12%
  addmm_100480x128x384: +0.91%
  addmm_6400x2048x512: +0.25%
  addmm_61568x1024x256: +0.08%
  addmm_1x768x768: +0.93%
  addmm_12544x384x384: +0.19%
  addmm_128x512x1000: +0.77%
  addmm_2048x128x128: +1.32%
  addmm_128x3072x1000: +0.24%
  addmm_7936x512x2048: +0.07%
  addmm_8192x512x2048: +0.33%
  addmm_64x1024x1000: +1.43%
  addmm_128x2304x1000: +0.01%
  addmm_32768x256x2: +0.75%
  addmm_64x384x1152: +0.79%
  addmm_64x640x1000: +0.01%
  addmm_100480x128x128: +0.87%
  addmm_1152x3072x768: +1.13%
  addmm_8192x256x2048: +1.40%
  addmm_4096x128x768: +0.01%
  addmm_128x2560x1000: +0.01%
  addmm_12544x2048x512: +0.43%
  addmm_200704x24x96: +0.14%
  addmm_8448x512x2048: +0.96%
  addmm_50176x256x1024: +0.62%
  addmm_4160x4096x1024: +0.22%
  addmm_4096x768x768: +0.32%
  addmm_220x2048x512: +0.56%
  addmm_8x2048x1000: +1.12%
  addmm_256x197951x512: +26.99%
  addmm_401536x64x192: +0.60%
  addmm_2040x2048x512: +0.47%
  addmm_512x1024x256: +1.32%
  addmm_128x4096x1000: +1.67%
  addmm_12672x768x768: +0.34%
  addmm_128x368x1000: +0.77%
  addmm_96x1280x1000: +0.01%
  addmm_12544x512x2048: +0.41%
  addmm_6272x320x1280: +0.76%
  addmm_12544x3072x768: +0.09%
  addmm_64x384x1000: +0.39%
mm improvements when best:
  mm_200704x128x512: +1.29%
  mm_663552x16x16: +0.80%
  mm_4096x768x768: +0.51%
  mm_131072x64x31: +0.24%
  mm_12544x1152x384: +0.11%
  mm_128x2048x2: +0.46%
  mm_262144x16x23: +0.62%
  mm_50176x576x192: +0.37%
  mm_131072x16x31: +0.26%
================================================================================
BENCHMARK ANALYSIS RESULTS
================================================================================

Operation: addmm
----------------------------------------
Total shapes analyzed: 247
Average Subgraph placement: 3.38
Median Subgraph placement: 2.0
Subgraph is best choice: 52/247 shapes (21.1%)
Average improvement when best: 1.15%
Median improvement when best: 0.58%
Largest improvement when best: +26.99%

Operation: bmm
----------------------------------------
Total shapes analyzed: 85
Average Subgraph placement: 24.00
Median Subgraph placement: 21.0
Subgraph is best choice: 0/85 shapes (0.0%)
Average improvement when best: N/A (never best)
Median improvement when best: N/A (never best)
Largest improvement when best: N/A (never best)

Operation: mm
----------------------------------------
Total shapes analyzed: 88
Average Subgraph placement: 15.08
Median Subgraph placement: 4.0
Subgraph is best choice: 9/88 shapes (10.2%)
Average improvement when best: 0.52%
Median improvement when best: 0.46%
Largest improvement when best: +1.29%

Results

The largest shape gain, 256,197951,512, seemed to be driven by a case where the extern kernel is way faster than the best triton configs on the recursive autotune:

addmm,Extern,extern_kernels.addmm,256,197951,512,0.38024500012397766
addmm,Triton,256,197951,512,32,256,16,2,2,4,2.005444049835205
addmm,Triton,256,197951,512,32,128,32,2,4,8,2.04189395904541
addmm,Triton,256,197951,512,64,128,16,2,4,8,2.1911399364471436
addmm,Triton,256,197951,512,64,128,32,2,4,8,2.496040105819702
addmm,Triton,256,197951,512,64,128,64,2,8,16,2.9306790828704834
addmm,Triton,256,197951,512,64,64,32,2,4,8,3.0347819328308105
...

Compared to the non-transposed autotune:

addmm,Subgraph,contiguous_addmm_1384,256,197951,512,0.5024129748344421
addmm,Extern,extern_kernels.addmm,256,197951,512,0.6881489753723145
addmm,Triton,256,197951,512,32,256,16,2,2,4,2.5115010738372803
addmm,Triton,256,197951,512,32,128,32,2,4,8,2.5167479515075684
addmm,Triton,256,197951,512,64,128,16,2,4,8,2.9507460594177246
addmm,Triton,256,197951,512,64,256,64,2,8,4,2.9673290252685547
addmm,Triton,256,197951,512,64,128,64,2,8,16,3.3906331062316895
addmm,Triton,256,197951,512,64,128,32,2,4,8,3.496859073638916

It seems to perform really well for high values of K vs N and M.
Testing this hypothesis with some custom shapes:

Parsed 64 unique shapes from benchmark output
addmm improvements when best:
  addmm_128x16384x128: +0.18%
  addmm_128x262144x256: +38.24%
  addmm_128x200000x512: +14.76%
  addmm_256x800000x128: +0.06%
  addmm_131072x128x256: +0.27%
  addmm_128x256x131072: +0.25%
  addmm_2048x200000x64: +12.45%
mm improvements when best:
  mm_128x16384x128: +0.18%
  mm_128x262144x256: +38.05%
  mm_128x200000x512: +9.47%
  mm_256x800000x128: +0.99%
  mm_512x6400000x256: +3.17%
  mm_524288x64x64: +0.29%
  mm_2048x200000x64: +11.19%
  mm_8192x1000000x256: +34.14%
  mm_128x4096x100000: +0.40%
  mm_128x3072x150000: +0.27%
================================================================================
BENCHMARK ANALYSIS RESULTS
================================================================================

Operation: addmm
----------------------------------------
Total shapes analyzed: 33
Average Subgraph placement: 4.39
Median Subgraph placement: 2.0
Subgraph is best choice: 7/33 shapes (21.2%)
Average improvement when best: 9.46%
Median improvement when best: 0.27%
Largest improvement when best: +38.24%

Operation: mm
----------------------------------------
Total shapes analyzed: 30
Average Subgraph placement: 7.63
Median Subgraph placement: 2.0
Subgraph is best choice: 10/30 shapes (33.3%)
Average improvement when best: 9.81%
Median improvement when best: 2.08%
Largest improvement when best: +38.05%

Conclusion

Contiguous Subgraph Decompositionseems worthwhile for mm and addmm, but not bmm, and has a very large improvment on low M, low N, and high K shapes.

Data gathering scripts:
https://gist.github.com/exclamaforte/4a896c064d301b27bf5ca0a4f8fc3866

Test Plan:

New unit tests.

Differential Revision: D80771648

cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben

pytorch-bot · 2025-08-22T03:25:23Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/161241

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit the bot commands wiki

Note: Links to docs will display an error until the docs builds have been completed.

✅ You can merge normally! (2 Unrelated Failures)

As of commit e05eae3 with merge base a99d8d3 ():

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

inductor / unit-test / inductor-halide-build / build (gh) (trunk failure)
undefined reference to NVPW_InitializeHost'`
inductor / unit-test / inductor-triton-cpu-build / build (gh) (trunk failure)
undefined reference to NVPW_InitializeHost'`

This comment was automatically generated by Dr. CI and updates every 15 minutes.

facebook-github-bot · 2025-08-22T03:25:30Z

This pull request was exported from Phabricator. Differential Revision: D80771648

facebook-github-bot · 2025-08-22T23:55:25Z

@exclamaforte has imported this pull request. If you are a Meta employee, you can view this in D80771648.

facebook-github-bot · 2025-08-26T22:22:28Z

@exclamaforte has imported this pull request. If you are a Meta employee, you can view this in D80771648.

eellison

looks good ! cc @PaulZhang12

torch/_inductor/kernel/mm.py

test/inductor/test_max_autotune.py

coconutruben · 2025-09-02T18:53:44Z

@exclamaforte can you do this in the same way we did #161026 ? this will allow us to integrate with everything else nicely

make a contigous_mm template
add a template heuristic for it, can just yield a single dict with no kwargs

exclamaforte · 2025-09-03T07:15:55Z

@pytorchbot merge

facebook-github-bot · 2025-09-03T07:16:11Z

@exclamaforte has imported this pull request. If you are a Meta employee, you can view this in D80771648.

pytorchmergebot · 2025-09-03T07:17:44Z

Merge failed

Reason: This PR has internal changes and must be landed via Phabricator! Please try reimporting/rexporting the PR!

Details for Dev Infra team

Raised by workflow job

pytorchmergebot · 2025-09-04T00:14:25Z

@exclamaforte your PR has been successfully reverted.

jeffdaily · 2025-09-04T00:16:04Z

See also https://hud.pytorch.org/failure?name=rocm-mi300%20%2F%20linux-noble-rocm-py3.12-mi300%20%2F%20test%20(default%2C%202%2C%206%2C%20linux.rocm.gpu.gfx942.1)&jobName=undefined&failureCaptures=inductor%2Ftest_max_autotune.py%3A%3ATestMaxAutotune%3A%3Atest_max_autotune_contiguous_transform_mm_float32_sizes2.

facebook-github-bot · 2025-09-04T01:21:30Z

@exclamaforte has imported this pull request. If you are a Meta employee, you can view this in D80771648.

exclamaforte · 2025-09-04T02:11:02Z

Verified that I fixed this test now on an AMD devserver

exclamaforte · 2025-09-04T02:11:08Z

@pytorchbot merge

pytorchmergebot · 2025-09-04T02:12:56Z

Merge failed

Reason: This PR has internal changes and must be landed via Phabricator! Please try reimporting/rexporting the PR!

Details for Dev Infra team

Raised by workflow job

facebook-github-bot · 2025-09-04T02:14:30Z

@exclamaforte has imported this pull request. If you are a Meta employee, you can view this in D80771648.

exclamaforte · 2025-09-04T02:15:42Z

@pytorchbot merge

pytorchmergebot · 2025-09-04T02:17:30Z

Merge started

Your 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

Advanced Debugging

Check the merge workflow status
here

## Summary Adds a subgraph decomposition for addmm and mm that performs well on large `K` compared to `M` and `N`, and functions well as an alternative to `split-k` on AMD (transposed only), which does not support AMD currently. ## Background On AMD (MI300x), for a matmul A * B, if B is non-contiguous, the resulting matmul is quite a bit slower. For example: ``` args[0]: TensorBox(StorageBox( InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1])) )) args[1]: TensorBox(StorageBox( InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[1, 178176])) )) ``` is a lot slower than: ``` args[0]: TensorBox(StorageBox( InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1])) )) args[1]: TensorBox(StorageBox( InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[6144, 1])) )) ``` This PR adds a subgraph decomposition to test out whether making B contiguous is faster than just using the normal kernels. ## Data I ran this on unique non-contiguous shapes from torchbench/huggingface and got these speedups: ``` Parsed 420 unique shapes from benchmark output addmm improvements when best: addmm_16448x512x2048: +0.14% addmm_128x2048x2048: +0.01% addmm_128x768x1000: +0.75% addmm_12672x3072x768: +1.08% addmm_512x768x32000: +0.62% addmm_12608x384x384: +0.00% addmm_4160x1024x4096: +0.90% addmm_16x768x2: +0.56% addmm_12608x3072x768: +0.09% addmm_64x4096x1000: +2.77% addmm_256x1024x512: +1.99% addmm_30x256x256: +1.12% addmm_100480x128x384: +0.91% addmm_6400x2048x512: +0.25% addmm_61568x1024x256: +0.08% addmm_1x768x768: +0.93% addmm_12544x384x384: +0.19% addmm_128x512x1000: +0.77% addmm_2048x128x128: +1.32% addmm_128x3072x1000: +0.24% addmm_7936x512x2048: +0.07% addmm_8192x512x2048: +0.33% addmm_64x1024x1000: +1.43% addmm_128x2304x1000: +0.01% addmm_32768x256x2: +0.75% addmm_64x384x1152: +0.79% addmm_64x640x1000: +0.01% addmm_100480x128x128: +0.87% addmm_1152x3072x768: +1.13% addmm_8192x256x2048: +1.40% addmm_4096x128x768: +0.01% addmm_128x2560x1000: +0.01% addmm_12544x2048x512: +0.43% addmm_200704x24x96: +0.14% addmm_8448x512x2048: +0.96% addmm_50176x256x1024: +0.62% addmm_4160x4096x1024: +0.22% addmm_4096x768x768: +0.32% addmm_220x2048x512: +0.56% addmm_8x2048x1000: +1.12% addmm_256x197951x512: +26.99% addmm_401536x64x192: +0.60% addmm_2040x2048x512: +0.47% addmm_512x1024x256: +1.32% addmm_128x4096x1000: +1.67% addmm_12672x768x768: +0.34% addmm_128x368x1000: +0.77% addmm_96x1280x1000: +0.01% addmm_12544x512x2048: +0.41% addmm_6272x320x1280: +0.76% addmm_12544x3072x768: +0.09% addmm_64x384x1000: +0.39% mm improvements when best: mm_200704x128x512: +1.29% mm_663552x16x16: +0.80% mm_4096x768x768: +0.51% mm_131072x64x31: +0.24% mm_12544x1152x384: +0.11% mm_128x2048x2: +0.46% mm_262144x16x23: +0.62% mm_50176x576x192: +0.37% mm_131072x16x31: +0.26% ================================================================================ BENCHMARK ANALYSIS RESULTS ================================================================================ Operation: addmm ---------------------------------------- Total shapes analyzed: 247 Average Subgraph placement: 3.38 Median Subgraph placement: 2.0 Subgraph is best choice: 52/247 shapes (21.1%) Average improvement when best: 1.15% Median improvement when best: 0.58% Largest improvement when best: +26.99% Operation: bmm ---------------------------------------- Total shapes analyzed: 85 Average Subgraph placement: 24.00 Median Subgraph placement: 21.0 Subgraph is best choice: 0/85 shapes (0.0%) Average improvement when best: N/A (never best) Median improvement when best: N/A (never best) Largest improvement when best: N/A (never best) Operation: mm ---------------------------------------- Total shapes analyzed: 88 Average Subgraph placement: 15.08 Median Subgraph placement: 4.0 Subgraph is best choice: 9/88 shapes (10.2%) Average improvement when best: 0.52% Median improvement when best: 0.46% Largest improvement when best: +1.29% ``` ## Results The largest shape gain, `256,197951,512`, seemed to be driven by a case where the extern kernel is way faster than the best triton configs on the recursive autotune: ``` addmm,Extern,extern_kernels.addmm,256,197951,512,0.38024500012397766 addmm,Triton,256,197951,512,32,256,16,2,2,4,2.005444049835205 addmm,Triton,256,197951,512,32,128,32,2,4,8,2.04189395904541 addmm,Triton,256,197951,512,64,128,16,2,4,8,2.1911399364471436 addmm,Triton,256,197951,512,64,128,32,2,4,8,2.496040105819702 addmm,Triton,256,197951,512,64,128,64,2,8,16,2.9306790828704834 addmm,Triton,256,197951,512,64,64,32,2,4,8,3.0347819328308105 ... ``` Compared to the non-transposed autotune: ``` addmm,Subgraph,contiguous_addmm_1384,256,197951,512,0.5024129748344421 addmm,Extern,extern_kernels.addmm,256,197951,512,0.6881489753723145 addmm,Triton,256,197951,512,32,256,16,2,2,4,2.5115010738372803 addmm,Triton,256,197951,512,32,128,32,2,4,8,2.5167479515075684 addmm,Triton,256,197951,512,64,128,16,2,4,8,2.9507460594177246 addmm,Triton,256,197951,512,64,256,64,2,8,4,2.9673290252685547 addmm,Triton,256,197951,512,64,128,64,2,8,16,3.3906331062316895 addmm,Triton,256,197951,512,64,128,32,2,4,8,3.496859073638916 ``` It seems to perform really well for high values of `K` vs `N` and `M`. Testing this hypothesis with some custom shapes: ``` Parsed 64 unique shapes from benchmark output addmm improvements when best: addmm_128x16384x128: +0.18% addmm_128x262144x256: +38.24% addmm_128x200000x512: +14.76% addmm_256x800000x128: +0.06% addmm_131072x128x256: +0.27% addmm_128x256x131072: +0.25% addmm_2048x200000x64: +12.45% mm improvements when best: mm_128x16384x128: +0.18% mm_128x262144x256: +38.05% mm_128x200000x512: +9.47% mm_256x800000x128: +0.99% mm_512x6400000x256: +3.17% mm_524288x64x64: +0.29% mm_2048x200000x64: +11.19% mm_8192x1000000x256: +34.14% mm_128x4096x100000: +0.40% mm_128x3072x150000: +0.27% ================================================================================ BENCHMARK ANALYSIS RESULTS ================================================================================ Operation: addmm ---------------------------------------- Total shapes analyzed: 33 Average Subgraph placement: 4.39 Median Subgraph placement: 2.0 Subgraph is best choice: 7/33 shapes (21.2%) Average improvement when best: 9.46% Median improvement when best: 0.27% Largest improvement when best: +38.24% Operation: mm ---------------------------------------- Total shapes analyzed: 30 Average Subgraph placement: 7.63 Median Subgraph placement: 2.0 Subgraph is best choice: 10/30 shapes (33.3%) Average improvement when best: 9.81% Median improvement when best: 2.08% Largest improvement when best: +38.05% ``` ## Conclusion Contiguous Subgraph Decompositionseems worthwhile for `mm` and `addmm`, but not `bmm`, and has a very large improvment on low `M`, low `N`, and high `K` shapes. Data gathering scripts: https://gist.github.com/exclamaforte/4a896c064d301b27bf5ca0a4f8fc3866 ## Test Plan: New unit tests. Differential Revision: D80771648 Pull Request resolved: pytorch#161241 Approved by: https://github.com/eellison

This reverts commit d647185. Reverted pytorch#161241 on behalf of https://github.com/jeffdaily due to breaks rocm mi300 tests ([comment](pytorch#161241 (comment)))

## Summary Adds a subgraph decomposition for addmm and mm that performs well on large `K` compared to `M` and `N`, and functions well as an alternative to `split-k` on AMD (transposed only), which does not support AMD currently. ## Background On AMD (MI300x), for a matmul A * B, if B is non-contiguous, the resulting matmul is quite a bit slower. For example: ``` args[0]: TensorBox(StorageBox( InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1])) )) args[1]: TensorBox(StorageBox( InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[1, 178176])) )) ``` is a lot slower than: ``` args[0]: TensorBox(StorageBox( InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1])) )) args[1]: TensorBox(StorageBox( InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[6144, 1])) )) ``` This PR adds a subgraph decomposition to test out whether making B contiguous is faster than just using the normal kernels. ## Data I ran this on unique non-contiguous shapes from torchbench/huggingface and got these speedups: ``` Parsed 420 unique shapes from benchmark output addmm improvements when best: addmm_16448x512x2048: +0.14% addmm_128x2048x2048: +0.01% addmm_128x768x1000: +0.75% addmm_12672x3072x768: +1.08% addmm_512x768x32000: +0.62% addmm_12608x384x384: +0.00% addmm_4160x1024x4096: +0.90% addmm_16x768x2: +0.56% addmm_12608x3072x768: +0.09% addmm_64x4096x1000: +2.77% addmm_256x1024x512: +1.99% addmm_30x256x256: +1.12% addmm_100480x128x384: +0.91% addmm_6400x2048x512: +0.25% addmm_61568x1024x256: +0.08% addmm_1x768x768: +0.93% addmm_12544x384x384: +0.19% addmm_128x512x1000: +0.77% addmm_2048x128x128: +1.32% addmm_128x3072x1000: +0.24% addmm_7936x512x2048: +0.07% addmm_8192x512x2048: +0.33% addmm_64x1024x1000: +1.43% addmm_128x2304x1000: +0.01% addmm_32768x256x2: +0.75% addmm_64x384x1152: +0.79% addmm_64x640x1000: +0.01% addmm_100480x128x128: +0.87% addmm_1152x3072x768: +1.13% addmm_8192x256x2048: +1.40% addmm_4096x128x768: +0.01% addmm_128x2560x1000: +0.01% addmm_12544x2048x512: +0.43% addmm_200704x24x96: +0.14% addmm_8448x512x2048: +0.96% addmm_50176x256x1024: +0.62% addmm_4160x4096x1024: +0.22% addmm_4096x768x768: +0.32% addmm_220x2048x512: +0.56% addmm_8x2048x1000: +1.12% addmm_256x197951x512: +26.99% addmm_401536x64x192: +0.60% addmm_2040x2048x512: +0.47% addmm_512x1024x256: +1.32% addmm_128x4096x1000: +1.67% addmm_12672x768x768: +0.34% addmm_128x368x1000: +0.77% addmm_96x1280x1000: +0.01% addmm_12544x512x2048: +0.41% addmm_6272x320x1280: +0.76% addmm_12544x3072x768: +0.09% addmm_64x384x1000: +0.39% mm improvements when best: mm_200704x128x512: +1.29% mm_663552x16x16: +0.80% mm_4096x768x768: +0.51% mm_131072x64x31: +0.24% mm_12544x1152x384: +0.11% mm_128x2048x2: +0.46% mm_262144x16x23: +0.62% mm_50176x576x192: +0.37% mm_131072x16x31: +0.26% ================================================================================ BENCHMARK ANALYSIS RESULTS ================================================================================ Operation: addmm ---------------------------------------- Total shapes analyzed: 247 Average Subgraph placement: 3.38 Median Subgraph placement: 2.0 Subgraph is best choice: 52/247 shapes (21.1%) Average improvement when best: 1.15% Median improvement when best: 0.58% Largest improvement when best: +26.99% Operation: bmm ---------------------------------------- Total shapes analyzed: 85 Average Subgraph placement: 24.00 Median Subgraph placement: 21.0 Subgraph is best choice: 0/85 shapes (0.0%) Average improvement when best: N/A (never best) Median improvement when best: N/A (never best) Largest improvement when best: N/A (never best) Operation: mm ---------------------------------------- Total shapes analyzed: 88 Average Subgraph placement: 15.08 Median Subgraph placement: 4.0 Subgraph is best choice: 9/88 shapes (10.2%) Average improvement when best: 0.52% Median improvement when best: 0.46% Largest improvement when best: +1.29% ``` ## Results The largest shape gain, `256,197951,512`, seemed to be driven by a case where the extern kernel is way faster than the best triton configs on the recursive autotune: ``` addmm,Extern,extern_kernels.addmm,256,197951,512,0.38024500012397766 addmm,Triton,256,197951,512,32,256,16,2,2,4,2.005444049835205 addmm,Triton,256,197951,512,32,128,32,2,4,8,2.04189395904541 addmm,Triton,256,197951,512,64,128,16,2,4,8,2.1911399364471436 addmm,Triton,256,197951,512,64,128,32,2,4,8,2.496040105819702 addmm,Triton,256,197951,512,64,128,64,2,8,16,2.9306790828704834 addmm,Triton,256,197951,512,64,64,32,2,4,8,3.0347819328308105 ... ``` Compared to the non-transposed autotune: ``` addmm,Subgraph,contiguous_addmm_1384,256,197951,512,0.5024129748344421 addmm,Extern,extern_kernels.addmm,256,197951,512,0.6881489753723145 addmm,Triton,256,197951,512,32,256,16,2,2,4,2.5115010738372803 addmm,Triton,256,197951,512,32,128,32,2,4,8,2.5167479515075684 addmm,Triton,256,197951,512,64,128,16,2,4,8,2.9507460594177246 addmm,Triton,256,197951,512,64,256,64,2,8,4,2.9673290252685547 addmm,Triton,256,197951,512,64,128,64,2,8,16,3.3906331062316895 addmm,Triton,256,197951,512,64,128,32,2,4,8,3.496859073638916 ``` It seems to perform really well for high values of `K` vs `N` and `M`. Testing this hypothesis with some custom shapes: ``` Parsed 64 unique shapes from benchmark output addmm improvements when best: addmm_128x16384x128: +0.18% addmm_128x262144x256: +38.24% addmm_128x200000x512: +14.76% addmm_256x800000x128: +0.06% addmm_131072x128x256: +0.27% addmm_128x256x131072: +0.25% addmm_2048x200000x64: +12.45% mm improvements when best: mm_128x16384x128: +0.18% mm_128x262144x256: +38.05% mm_128x200000x512: +9.47% mm_256x800000x128: +0.99% mm_512x6400000x256: +3.17% mm_524288x64x64: +0.29% mm_2048x200000x64: +11.19% mm_8192x1000000x256: +34.14% mm_128x4096x100000: +0.40% mm_128x3072x150000: +0.27% ================================================================================ BENCHMARK ANALYSIS RESULTS ================================================================================ Operation: addmm ---------------------------------------- Total shapes analyzed: 33 Average Subgraph placement: 4.39 Median Subgraph placement: 2.0 Subgraph is best choice: 7/33 shapes (21.2%) Average improvement when best: 9.46% Median improvement when best: 0.27% Largest improvement when best: +38.24% Operation: mm ---------------------------------------- Total shapes analyzed: 30 Average Subgraph placement: 7.63 Median Subgraph placement: 2.0 Subgraph is best choice: 10/30 shapes (33.3%) Average improvement when best: 9.81% Median improvement when best: 2.08% Largest improvement when best: +38.05% ``` ## Conclusion Contiguous Subgraph Decompositionseems worthwhile for `mm` and `addmm`, but not `bmm`, and has a very large improvment on low `M`, low `N`, and high `K` shapes. Data gathering scripts: https://gist.github.com/exclamaforte/4a896c064d301b27bf5ca0a4f8fc3866 ## Test Plan: New unit tests. Differential Revision: D80771648 Pull Request resolved: pytorch#161241 Approved by: https://github.com/eellison

This reverts commit d647185. Reverted pytorch#161241 on behalf of https://github.com/jeffdaily due to breaks rocm mi300 tests ([comment](pytorch#161241 (comment)))

## Summary Adds a subgraph decomposition for addmm and mm that performs well on large `K` compared to `M` and `N`, and functions well as an alternative to `split-k` on AMD (transposed only), which does not support AMD currently. ## Background On AMD (MI300x), for a matmul A * B, if B is non-contiguous, the resulting matmul is quite a bit slower. For example: ``` args[0]: TensorBox(StorageBox( InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1])) )) args[1]: TensorBox(StorageBox( InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[1, 178176])) )) ``` is a lot slower than: ``` args[0]: TensorBox(StorageBox( InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1])) )) args[1]: TensorBox(StorageBox( InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[6144, 1])) )) ``` This PR adds a subgraph decomposition to test out whether making B contiguous is faster than just using the normal kernels. ## Data I ran this on unique non-contiguous shapes from torchbench/huggingface and got these speedups: ``` Parsed 420 unique shapes from benchmark output addmm improvements when best: addmm_16448x512x2048: +0.14% addmm_128x2048x2048: +0.01% addmm_128x768x1000: +0.75% addmm_12672x3072x768: +1.08% addmm_512x768x32000: +0.62% addmm_12608x384x384: +0.00% addmm_4160x1024x4096: +0.90% addmm_16x768x2: +0.56% addmm_12608x3072x768: +0.09% addmm_64x4096x1000: +2.77% addmm_256x1024x512: +1.99% addmm_30x256x256: +1.12% addmm_100480x128x384: +0.91% addmm_6400x2048x512: +0.25% addmm_61568x1024x256: +0.08% addmm_1x768x768: +0.93% addmm_12544x384x384: +0.19% addmm_128x512x1000: +0.77% addmm_2048x128x128: +1.32% addmm_128x3072x1000: +0.24% addmm_7936x512x2048: +0.07% addmm_8192x512x2048: +0.33% addmm_64x1024x1000: +1.43% addmm_128x2304x1000: +0.01% addmm_32768x256x2: +0.75% addmm_64x384x1152: +0.79% addmm_64x640x1000: +0.01% addmm_100480x128x128: +0.87% addmm_1152x3072x768: +1.13% addmm_8192x256x2048: +1.40% addmm_4096x128x768: +0.01% addmm_128x2560x1000: +0.01% addmm_12544x2048x512: +0.43% addmm_200704x24x96: +0.14% addmm_8448x512x2048: +0.96% addmm_50176x256x1024: +0.62% addmm_4160x4096x1024: +0.22% addmm_4096x768x768: +0.32% addmm_220x2048x512: +0.56% addmm_8x2048x1000: +1.12% addmm_256x197951x512: +26.99% addmm_401536x64x192: +0.60% addmm_2040x2048x512: +0.47% addmm_512x1024x256: +1.32% addmm_128x4096x1000: +1.67% addmm_12672x768x768: +0.34% addmm_128x368x1000: +0.77% addmm_96x1280x1000: +0.01% addmm_12544x512x2048: +0.41% addmm_6272x320x1280: +0.76% addmm_12544x3072x768: +0.09% addmm_64x384x1000: +0.39% mm improvements when best: mm_200704x128x512: +1.29% mm_663552x16x16: +0.80% mm_4096x768x768: +0.51% mm_131072x64x31: +0.24% mm_12544x1152x384: +0.11% mm_128x2048x2: +0.46% mm_262144x16x23: +0.62% mm_50176x576x192: +0.37% mm_131072x16x31: +0.26% ================================================================================ BENCHMARK ANALYSIS RESULTS ================================================================================ Operation: addmm ---------------------------------------- Total shapes analyzed: 247 Average Subgraph placement: 3.38 Median Subgraph placement: 2.0 Subgraph is best choice: 52/247 shapes (21.1%) Average improvement when best: 1.15% Median improvement when best: 0.58% Largest improvement when best: +26.99% Operation: bmm ---------------------------------------- Total shapes analyzed: 85 Average Subgraph placement: 24.00 Median Subgraph placement: 21.0 Subgraph is best choice: 0/85 shapes (0.0%) Average improvement when best: N/A (never best) Median improvement when best: N/A (never best) Largest improvement when best: N/A (never best) Operation: mm ---------------------------------------- Total shapes analyzed: 88 Average Subgraph placement: 15.08 Median Subgraph placement: 4.0 Subgraph is best choice: 9/88 shapes (10.2%) Average improvement when best: 0.52% Median improvement when best: 0.46% Largest improvement when best: +1.29% ``` ## Results The largest shape gain, `256,197951,512`, seemed to be driven by a case where the extern kernel is way faster than the best triton configs on the recursive autotune: ``` addmm,Extern,extern_kernels.addmm,256,197951,512,0.38024500012397766 addmm,Triton,256,197951,512,32,256,16,2,2,4,2.005444049835205 addmm,Triton,256,197951,512,32,128,32,2,4,8,2.04189395904541 addmm,Triton,256,197951,512,64,128,16,2,4,8,2.1911399364471436 addmm,Triton,256,197951,512,64,128,32,2,4,8,2.496040105819702 addmm,Triton,256,197951,512,64,128,64,2,8,16,2.9306790828704834 addmm,Triton,256,197951,512,64,64,32,2,4,8,3.0347819328308105 ... ``` Compared to the non-transposed autotune: ``` addmm,Subgraph,contiguous_addmm_1384,256,197951,512,0.5024129748344421 addmm,Extern,extern_kernels.addmm,256,197951,512,0.6881489753723145 addmm,Triton,256,197951,512,32,256,16,2,2,4,2.5115010738372803 addmm,Triton,256,197951,512,32,128,32,2,4,8,2.5167479515075684 addmm,Triton,256,197951,512,64,128,16,2,4,8,2.9507460594177246 addmm,Triton,256,197951,512,64,256,64,2,8,4,2.9673290252685547 addmm,Triton,256,197951,512,64,128,64,2,8,16,3.3906331062316895 addmm,Triton,256,197951,512,64,128,32,2,4,8,3.496859073638916 ``` It seems to perform really well for high values of `K` vs `N` and `M`. Testing this hypothesis with some custom shapes: ``` Parsed 64 unique shapes from benchmark output addmm improvements when best: addmm_128x16384x128: +0.18% addmm_128x262144x256: +38.24% addmm_128x200000x512: +14.76% addmm_256x800000x128: +0.06% addmm_131072x128x256: +0.27% addmm_128x256x131072: +0.25% addmm_2048x200000x64: +12.45% mm improvements when best: mm_128x16384x128: +0.18% mm_128x262144x256: +38.05% mm_128x200000x512: +9.47% mm_256x800000x128: +0.99% mm_512x6400000x256: +3.17% mm_524288x64x64: +0.29% mm_2048x200000x64: +11.19% mm_8192x1000000x256: +34.14% mm_128x4096x100000: +0.40% mm_128x3072x150000: +0.27% ================================================================================ BENCHMARK ANALYSIS RESULTS ================================================================================ Operation: addmm ---------------------------------------- Total shapes analyzed: 33 Average Subgraph placement: 4.39 Median Subgraph placement: 2.0 Subgraph is best choice: 7/33 shapes (21.2%) Average improvement when best: 9.46% Median improvement when best: 0.27% Largest improvement when best: +38.24% Operation: mm ---------------------------------------- Total shapes analyzed: 30 Average Subgraph placement: 7.63 Median Subgraph placement: 2.0 Subgraph is best choice: 10/30 shapes (33.3%) Average improvement when best: 9.81% Median improvement when best: 2.08% Largest improvement when best: +38.05% ``` ## Conclusion Contiguous Subgraph Decompositionseems worthwhile for `mm` and `addmm`, but not `bmm`, and has a very large improvment on low `M`, low `N`, and high `K` shapes. Data gathering scripts: https://gist.github.com/exclamaforte/4a896c064d301b27bf5ca0a4f8fc3866 ## Test Plan: New unit tests. Differential Revision: D80771648 Pull Request resolved: pytorch#161241 Approved by: https://github.com/eellison

This reverts commit d647185. Reverted pytorch#161241 on behalf of https://github.com/jeffdaily due to breaks rocm mi300 tests ([comment](pytorch#161241 (comment)))

## Summary Adds a subgraph decomposition for addmm and mm that performs well on large `K` compared to `M` and `N`, and functions well as an alternative to `split-k` on AMD (transposed only), which does not support AMD currently. ## Background On AMD (MI300x), for a matmul A * B, if B is non-contiguous, the resulting matmul is quite a bit slower. For example: ``` args[0]: TensorBox(StorageBox( InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1])) )) args[1]: TensorBox(StorageBox( InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[1, 178176])) )) ``` is a lot slower than: ``` args[0]: TensorBox(StorageBox( InputBuffer(name='arg0_1', layout=FixedLayout('cuda:0', torch.float16, size=[1024, 178176], stride=[178176, 1])) )) args[1]: TensorBox(StorageBox( InputBuffer(name='arg1_1', layout=FixedLayout('cuda:0', torch.float16, size=[178176, 6144], stride=[6144, 1])) )) ``` This PR adds a subgraph decomposition to test out whether making B contiguous is faster than just using the normal kernels. ## Data I ran this on unique non-contiguous shapes from torchbench/huggingface and got these speedups: ``` Parsed 420 unique shapes from benchmark output addmm improvements when best: addmm_16448x512x2048: +0.14% addmm_128x2048x2048: +0.01% addmm_128x768x1000: +0.75% addmm_12672x3072x768: +1.08% addmm_512x768x32000: +0.62% addmm_12608x384x384: +0.00% addmm_4160x1024x4096: +0.90% addmm_16x768x2: +0.56% addmm_12608x3072x768: +0.09% addmm_64x4096x1000: +2.77% addmm_256x1024x512: +1.99% addmm_30x256x256: +1.12% addmm_100480x128x384: +0.91% addmm_6400x2048x512: +0.25% addmm_61568x1024x256: +0.08% addmm_1x768x768: +0.93% addmm_12544x384x384: +0.19% addmm_128x512x1000: +0.77% addmm_2048x128x128: +1.32% addmm_128x3072x1000: +0.24% addmm_7936x512x2048: +0.07% addmm_8192x512x2048: +0.33% addmm_64x1024x1000: +1.43% addmm_128x2304x1000: +0.01% addmm_32768x256x2: +0.75% addmm_64x384x1152: +0.79% addmm_64x640x1000: +0.01% addmm_100480x128x128: +0.87% addmm_1152x3072x768: +1.13% addmm_8192x256x2048: +1.40% addmm_4096x128x768: +0.01% addmm_128x2560x1000: +0.01% addmm_12544x2048x512: +0.43% addmm_200704x24x96: +0.14% addmm_8448x512x2048: +0.96% addmm_50176x256x1024: +0.62% addmm_4160x4096x1024: +0.22% addmm_4096x768x768: +0.32% addmm_220x2048x512: +0.56% addmm_8x2048x1000: +1.12% addmm_256x197951x512: +26.99% addmm_401536x64x192: +0.60% addmm_2040x2048x512: +0.47% addmm_512x1024x256: +1.32% addmm_128x4096x1000: +1.67% addmm_12672x768x768: +0.34% addmm_128x368x1000: +0.77% addmm_96x1280x1000: +0.01% addmm_12544x512x2048: +0.41% addmm_6272x320x1280: +0.76% addmm_12544x3072x768: +0.09% addmm_64x384x1000: +0.39% mm improvements when best: mm_200704x128x512: +1.29% mm_663552x16x16: +0.80% mm_4096x768x768: +0.51% mm_131072x64x31: +0.24% mm_12544x1152x384: +0.11% mm_128x2048x2: +0.46% mm_262144x16x23: +0.62% mm_50176x576x192: +0.37% mm_131072x16x31: +0.26% ================================================================================ BENCHMARK ANALYSIS RESULTS ================================================================================ Operation: addmm ---------------------------------------- Total shapes analyzed: 247 Average Subgraph placement: 3.38 Median Subgraph placement: 2.0 Subgraph is best choice: 52/247 shapes (21.1%) Average improvement when best: 1.15% Median improvement when best: 0.58% Largest improvement when best: +26.99% Operation: bmm ---------------------------------------- Total shapes analyzed: 85 Average Subgraph placement: 24.00 Median Subgraph placement: 21.0 Subgraph is best choice: 0/85 shapes (0.0%) Average improvement when best: N/A (never best) Median improvement when best: N/A (never best) Largest improvement when best: N/A (never best) Operation: mm ---------------------------------------- Total shapes analyzed: 88 Average Subgraph placement: 15.08 Median Subgraph placement: 4.0 Subgraph is best choice: 9/88 shapes (10.2%) Average improvement when best: 0.52% Median improvement when best: 0.46% Largest improvement when best: +1.29% ``` ## Results The largest shape gain, `256,197951,512`, seemed to be driven by a case where the extern kernel is way faster than the best triton configs on the recursive autotune: ``` addmm,Extern,extern_kernels.addmm,256,197951,512,0.38024500012397766 addmm,Triton,256,197951,512,32,256,16,2,2,4,2.005444049835205 addmm,Triton,256,197951,512,32,128,32,2,4,8,2.04189395904541 addmm,Triton,256,197951,512,64,128,16,2,4,8,2.1911399364471436 addmm,Triton,256,197951,512,64,128,32,2,4,8,2.496040105819702 addmm,Triton,256,197951,512,64,128,64,2,8,16,2.9306790828704834 addmm,Triton,256,197951,512,64,64,32,2,4,8,3.0347819328308105 ... ``` Compared to the non-transposed autotune: ``` addmm,Subgraph,contiguous_addmm_1384,256,197951,512,0.5024129748344421 addmm,Extern,extern_kernels.addmm,256,197951,512,0.6881489753723145 addmm,Triton,256,197951,512,32,256,16,2,2,4,2.5115010738372803 addmm,Triton,256,197951,512,32,128,32,2,4,8,2.5167479515075684 addmm,Triton,256,197951,512,64,128,16,2,4,8,2.9507460594177246 addmm,Triton,256,197951,512,64,256,64,2,8,4,2.9673290252685547 addmm,Triton,256,197951,512,64,128,64,2,8,16,3.3906331062316895 addmm,Triton,256,197951,512,64,128,32,2,4,8,3.496859073638916 ``` It seems to perform really well for high values of `K` vs `N` and `M`. Testing this hypothesis with some custom shapes: ``` Parsed 64 unique shapes from benchmark output addmm improvements when best: addmm_128x16384x128: +0.18% addmm_128x262144x256: +38.24% addmm_128x200000x512: +14.76% addmm_256x800000x128: +0.06% addmm_131072x128x256: +0.27% addmm_128x256x131072: +0.25% addmm_2048x200000x64: +12.45% mm improvements when best: mm_128x16384x128: +0.18% mm_128x262144x256: +38.05% mm_128x200000x512: +9.47% mm_256x800000x128: +0.99% mm_512x6400000x256: +3.17% mm_524288x64x64: +0.29% mm_2048x200000x64: +11.19% mm_8192x1000000x256: +34.14% mm_128x4096x100000: +0.40% mm_128x3072x150000: +0.27% ================================================================================ BENCHMARK ANALYSIS RESULTS ================================================================================ Operation: addmm ---------------------------------------- Total shapes analyzed: 33 Average Subgraph placement: 4.39 Median Subgraph placement: 2.0 Subgraph is best choice: 7/33 shapes (21.2%) Average improvement when best: 9.46% Median improvement when best: 0.27% Largest improvement when best: +38.24% Operation: mm ---------------------------------------- Total shapes analyzed: 30 Average Subgraph placement: 7.63 Median Subgraph placement: 2.0 Subgraph is best choice: 10/30 shapes (33.3%) Average improvement when best: 9.81% Median improvement when best: 2.08% Largest improvement when best: +38.05% ``` ## Conclusion Contiguous Subgraph Decompositionseems worthwhile for `mm` and `addmm`, but not `bmm`, and has a very large improvment on low `M`, low `N`, and high `K` shapes. Data gathering scripts: https://gist.github.com/exclamaforte/4a896c064d301b27bf5ca0a4f8fc3866 ## Test Plan: New unit tests. Differential Revision: D80771648 Pull Request resolved: pytorch#161241 Approved by: https://github.com/eellison

pytorch-bot bot added ciflow/inductor module: inductor labels Aug 22, 2025

facebook-github-bot added the fb-exported label Aug 22, 2025

exclamaforte added the topic: not user facing topic category label Aug 22, 2025

pytorch-bot bot added the ciflow/trunk Trigger trunk jobs on your pull request label Aug 22, 2025

exclamaforte force-pushed the export-D80771648 branch 3 times, most recently from 4a42dca to 1bb3352 Compare August 26, 2025 17:25

exclamaforte requested a review from eellison August 26, 2025 22:18

exclamaforte changed the title ~~Contiguous transpose subgraph decomposition~~ Contiguous subgraph decomposition Aug 26, 2025

exclamaforte requested review from jansel and removed request for eellison and jansel August 27, 2025 04:36

exclamaforte force-pushed the export-D80771648 branch from 7c850eb to 911a8b0 Compare August 27, 2025 19:28

exclamaforte requested review from BoyuanFeng and eellison September 1, 2025 20:44

eellison approved these changes Sep 2, 2025

View reviewed changes

torch/_inductor/kernel/mm.py Outdated Show resolved Hide resolved

test/inductor/test_max_autotune.py Outdated Show resolved Hide resolved

test/inductor/test_max_autotune.py Show resolved Hide resolved

contiguous subgraph decomposition

6fd9b1b

exclamaforte force-pushed the export-D80771648 branch from d79d6ff to 6fd9b1b Compare September 2, 2025 18:04

exclamaforte added 2 commits September 3, 2025 00:11

update to contiguous mm

05d86f7

lint

39400c8

pytorchmergebot added the merging label Sep 3, 2025

pytorchmergebot added Reverted ci-no-td Do not run TD on this PR labels Sep 4, 2025

pytorchmergebot reopened this Sep 4, 2025

jeffdaily added the ciflow/rocm-mi300 Trigger "default" config CI on ROCm MI300 label Sep 4, 2025

fix

e05eae3

pytorchmergebot added the merging label Sep 4, 2025

pytorchmergebot removed the merging label Sep 4, 2025

pytorchmergebot added the merging label Sep 4, 2025

pytorchmergebot closed this in 66f3b4a Sep 4, 2025

pytorchmergebot removed the merging label Sep 4, 2025

github-actions bot deleted the export-D80771648 branch October 5, 2025 02:17

Conversation

exclamaforte commented Aug 22, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Background

Data

Results

Conclusion

Test Plan:

Uh oh!

pytorch-bot bot commented Aug 22, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/161241

✅ You can merge normally! (2 Unrelated Failures)

Uh oh!

facebook-github-bot commented Aug 22, 2025

Uh oh!

facebook-github-bot commented Aug 22, 2025

Uh oh!

facebook-github-bot commented Aug 26, 2025

Uh oh!

eellison left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

coconutruben commented Sep 2, 2025

Uh oh!

exclamaforte commented Sep 3, 2025

Uh oh!

facebook-github-bot commented Sep 3, 2025

Uh oh!

pytorchmergebot commented Sep 3, 2025

Merge failed

Uh oh!

pytorchmergebot commented Sep 4, 2025

Uh oh!

jeffdaily commented Sep 4, 2025

Uh oh!

facebook-github-bot commented Sep 4, 2025

Uh oh!

exclamaforte commented Sep 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

exclamaforte commented Sep 4, 2025

Uh oh!

pytorchmergebot commented Sep 4, 2025

Merge failed

Uh oh!

facebook-github-bot commented Sep 4, 2025

Uh oh!

exclamaforte commented Sep 4, 2025

Uh oh!

pytorchmergebot commented Sep 4, 2025

Merge started

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

6 participants

exclamaforte commented Aug 22, 2025 •

edited

Loading

pytorch-bot bot commented Aug 22, 2025 •

edited

Loading

exclamaforte commented Sep 4, 2025 •

edited

Loading