[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

Inductor has no transposed-convolution template on CUDA. The backend rewrites conv_transpose1d and conv_transpose2d as convolutions, but aten.convolution with transposed=True (from to_edge) and conv_transpose3d failed to lower.

A transposed convolution is the same GEMM over its output pixels: output index o reads, at tap k, input (o + padding_before - k) / stride when that is an integer inside the input, and the [C_in, C_out, ...] weight is read with its first two strides swapped. One constexpr adds this gather to the implicit-GEMM kernel (with pixel tiles the image base is a scalar, so the gather adds it without indexing it, which Triton 3.9 rejects); schedules, configs, store and output layouts are shared. supports() takes dense transposed 1-D, 2-D and 3-D calls with any stride, output_padding and crop, including the asymmetric crop only the canonical op expresses. The router takes them only where the regular lowering cannot compile them, and worth_routing refuses them otherwise. The benchmark's shapes suite gains transposed 1-D upsampling (kernel 4, stride 2), timed against the backend's rewrite.

test_conv_implicit_gemm.py runs every schedule on transposed geometries (stride, output_padding, asymmetric crop, tap-major K order) in every dtype and input layout, and lowers transposed 1-D and 3-D convolutions through both APIs against eager. test_conv_routing.py checks transposed 1-D, 2-D and 3-D routing and keeps dilated and grouped ones.

This diff was authored with Claude Code.

Differential Revision: D129825956

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/25091

• 📄 Preview Python docs built from this PR

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

❌ 1 Cancelled Job, 1 Pending, 1 Unrelated Failure

As of commit 9b4331e with merge base dea39f0 ():

CANCELLED JOB - The following job was cancelled. Please retry:

• Test CUDA Builds / unittest-cuda (system) / linux-job (gh)
  ##[error]received SIGINT

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.