Conversation
test_eager_correctness, test_single_expert, and test_batched_correctness used 5% relative tolerance for INT4 kernel-vs-dequant comparison. Tighten to 2% to match the e2e runner bar (fe71bd4).
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18934
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ You can merge normally! (2 Unrelated Failures)As of commit 75a2348 with merge base ec8d70b ( BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
digantdesai
changed the title
This PR needs a
|
There was a problem hiding this comment.
Pull request overview
Removes CUDA AOTI kernel/op benchmarking scripts from the repo, and also makes fused MoE correctness tests stricter to catch larger numeric discrepancies.
Changes:
- Delete Triton SDPA benchmark script (
benchmark_sdpa.py). - Delete fused MoE benchmark script (
benchmark_moe.py). - Tighten fused MoE test relative-diff tolerance from
0.05to0.02.
Reviewed changes
Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
| backends/cuda/tests/test_fused_moe.py | Tightens numerical correctness thresholds for fused MoE kernels. |
| backends/cuda/benchmarks/benchmark_sdpa.py | Removed SDPA benchmarking script (Triton vs PyTorch SDPA backends). |
| backends/cuda/benchmarks/benchmark_moe.py | Removed fused MoE benchmarking script (Triton vs eager/compile baselines). |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| self.assertLess( | ||
| rel, | ||
| 0.05, | ||
| 0.02, | ||
| f"{desc}: relative diff {rel:.4f} (abs {diff:.6f})", |
There was a problem hiding this comment.
maybe we should create a global const for the threshold
No description provided.