[Gasoonjia]

unittest-cuda has hit its 90 min timeout on every main run since a04a9f9 (#24004). That commit made the SDPA autotune tests about 6.5 min slower (test_triton_sdpa.py went from 10.4 to 16.9 min). The job already took 83–89 min before it, so the extra time pushed it over the limit. Example run: https://github.com/pytorch/executorch/actions/runs/39464763398/job/118773442760

This PR splits the job into a two-shard matrix:

• unittest-cuda (kernel): Triton kernel and autotune tests. That covers SDPA, split-K SDPA, SDPA NaN, tq4 SDPA, int4 matmul, int5 large GEMM, quantized GEMM family, fused MoE, fused deltanet decode, chunk gated delta rule, topk, autotune inputs and launch params, and the Triton mutation cache. About 42 min of pytest.
• unittest-cuda (system): everything else under backends/cuda/tests and backends/cuda/passes/tests. That covers the lowering passes (off-graph KV, DFlash off-graph KV, SDPA split-K replacement), quantized op dispatch, and export, partitioner and runtime tests. This shard also runs the pybindings device tests, the C++ shim tests, and the Qwen 3.5 MoE and Gemma 4 31B model tests. About 28 min of pytest.

Kernel tests are listed explicitly in KERNEL_TESTS. The system shard runs the test directories and passes --ignore for the kernel files and for the quantized GEMM files that unittest-cuda-quantized-gemm already runs, so any new test file goes to the system shard by default. Both shards still do the full install and build, including executor_runner, because both have e2e tests that need it. Both pytest calls now pass --durations=20.

Test plan:

• Checked that the YAML parses, that every listed test file exists, and that the kernel shard (14 files), system shard (19), and quantized GEMM job (7) together cover all 40 backends/cuda/tests/test_*.py files with no overlap.
• Ran bash -n on the script for each shard.
• python -m unittest .ci/scripts/tests/test_muse_glimmer_batching_ci.py passes. That test parses cuda.yml.
• CI on this PR: both unittest-cuda shards should finish well under 90 min.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24047

• 📄 Preview Python docs built from this PR

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

⏳ 8 Pending, 1 Unrelated Failure

As of commit f800e4b with merge base 6368414 ():

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

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

• Test CUDA Builds / unittest-cuda-muse-glimmer-batching / linux-job (gh) (trunk failure)
  [ FAILED ] CudaDFlashNative.FinalWholeBlockFitsExactlyAndPreservesAllPrefillRows

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

[linux-foundation-easycla]

The committers listed above are authorized under a signed CLA.

• ✅ login: Gasoonjia / name: gasoonjia (491e1ca)

[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.