[Gasoonjia]

Stack from ghstack (oldest at bottom):

• [executorch][cuda] Run transposed convolutions on the implicit-GEMM template #25091
• [executorch][cuda] Test 3-D convolutions on the implicit-GEMM template #25090
• [executorch][cuda] Run 1-D convolutions on the implicit-GEMM template #25089
• [executorch][cuda] Converted convolutions follow the output layout the graph recorded #25088
• -> [executorch][cuda] Route 3-input-channel window and kernel-equals-stride convolutions to the implicit-GEMM template #25087
• [executorch][cuda] Add an implicit-GEMM template for dense triton::convolution #25086
• [executorch][cuda] Route 2-D depthwise and few-channel grouped convolutions to a Triton direct-stencil kernel #25085
• [executorch][cuda] Add triton::convolution and the conv routing framework #25084
• [executorch][cuda] Add planned replacement to the Triton replacement pass #25083
• [executorch][cuda] Deprecate triton_kernel_mode=OFF #25082

On convolutions with 3 input channels, Inductor's Triton conv template spends most of each channel block on empty lanes. The implicit-GEMM template walks the 3 channels without them, but no call is routed to it.

worth_routing now takes 3-channel calls whose (kernel, stride, symmetric padding) is in ROUTED_FAMILIES: 3x3, 5x5 and 7x7 windows with stride 1, 2 or 4 and padding 0 or half the kernel, 2-D convolutions with kernel == stride from 2 to 32, and 3-D ones with kernel == stride (1,4,4), (2,14,14) and (2,16,16). It bounds batch, extent, input pixels, output pixels, output channels and output elements, and requires a dense NCHW or channels-last input, a contiguous weight, and tensor cores for fp16, bf16 and TF32. Everything else, 1x1 convolutions included, keeps the regular lowering with a recorded reason. benchmarks/conv/benchmark_implicit_gemm.py measures boundary_calls(), every family geometry at every corner of these limits: all 224 are significantly faster than Inductor on an A100 in fp32, fp16, bf16 and TF32, with NCHW and channels-last inputs. The benchmark also has shapes and envelope suites.

test_conv_routing.py derives its range tests from boundary_calls(): every corner routed, one step past each limit and unbenchmarked layouts kept. test_conv_implicit_gemm.py lowers routed models and runs them against eager.

This diff was authored with Claude Code.

Differential Revision: D129645242

NOTE FOR REVIEWERS: This PR has internal Meta-specific changes or comments, please review them on Phabricator!

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/25087

• 📄 Preview Python docs built from this PR

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

⏳ 1 Pending, 1 Unrelated Failure

As of commit 32824bb with merge base dea39f0 ():

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

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

• Test ROCm AOTI / test-voxtral-realtime-rocm-gfx950-rocm7.2 / linux-job (gh) (trunk failure)
  ##[error]The operation was canceled.

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

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.