Skip to content

[https://nvbugs/6966263][test] Re-waive the flaky CuTe DSL MoE NVFP4 act-fusion test at its moe/kernels path - #20800

Merged
lori-ren merged 2 commits into
NVIDIA:mainfrom
lori-ren:test/rewaive-cute-dsl-moe-nvfp4
Oct 23, 2026
Merged

lori-ren merged 2 commits into
NVIDIA:mainfrom
lori-ren:test/rewaive-cute-dsl-moe-nvfp4

Conversation

@lori-ren

@lori-ren lori-ren commented Oct 22, 2026 •

Copy link
Copy Markdown
Contributor

Dev Engineer Review

The waiver now targets test_cute_dsl_moe.py::test_nvfp4_gather_grouped_gemm_act_fusion_blackwell instead of the unittest/_torch/thop/parallel directory. This limits the waiver to the flaky test and leaves other tests eligible to run. Review finding counts are unavailable.

QA Engineer Review

Only tests/integration/test_lists/waives.txt changed. It removes the directory-level waiver and adds a waiver for the Blackwell Cute DSL MoE grouped-GEMM fusion test under NVBUG 6966263. The supplied CI report says the pipeline succeeded after automatic test retries. Coverage verdict: sufficient.

Per-File QA Perspective

  • tests/integration/test_lists/waives.txt: The waiver targets the specific test at its moved path. Other tests are no longer excluded by this entry.

Description

#20735 waived unittest/_torch/thop/parallel under nvbugs/6966263, but #20658 moved that directory's tests to unittest/_torch/moe/kernels/parallel, so the waive no longer matches anything and the flaky case runs again on main.

This replaces it with a function-level waive of test_cute_dsl_moe.py::test_nvfp4_gather_grouped_gemm_act_fusion_blackwell, the case the bug tracks. The old entry skipped the whole directory, including test_moe_densegemm.py and the other CuTe DSL MoE tests; with the waive narrowed to this case, the rest of the directory runs again.

Test Coverage

Waive change only; the waives.txt pre-commit checks (duplicates, entries exist in source) pass.

PR Checklist

Please review the following before submitting your PR:

  • PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

  • PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

  • Test cases are provided for new code paths (see test instructions)

  • If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

  • Any new dependencies have been scanned for license and vulnerabilities

  • CODEOWNERS updated if ownership changes

  • Documentation updated as needed

  • Update tava architecture diagram if there is a significant design change in PR.

  • The reviewers assigned automatically/manually are appropriate for the PR.

  • Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

…act-fusion test at its moe/kernels path

The directory waive from NVIDIA#20735 named the pre-move unittest/_torch/thop/parallel path and no longer matches; waive only the tracked test.

Signed-off-by: Lori Ren <lorir@nvidia.com>
@lori-ren

Copy link
Copy Markdown
Contributor Author

/bot run

@github-actions

Copy link
Copy Markdown

Automatically added "ci: full pre-merge approved" because this PR has satisfied the required GitHub review approvals. Unresolved review conversations and other required checks remain independent merge requirements.

@lori-ren
lori-ren disabled auto-merge October 22, 2026 05:34
@lori-ren
lori-ren enabled auto-merge (squash) October 22, 2026 05:35
@coderabbitai

coderabbitai Bot commented Oct 22, 2026 •

Copy link
Copy Markdown
Contributor

Review in Change Stack →

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info
⚙️ Run configuration
📥 Commits

Reviewing files that changed from the base of the PR and between 979f086 and ab85d2b.

📒 Files selected for processing: 1
  • tests/integration/test_lists/waives.txt

Included review availability: This review used your included allowance. Your plan provides up to 12 included reviews per hour; 11 remain after this review.

Walkthrough

The integration test waiver list adds a skip entry for a Blackwell Cute DSL MoE grouped-GEMM fusion test and removes the skip entry for unittest/_torch/thop/parallel. Both entries reference issue 6966263.

Changes

Integration test waiver updates

Layer / File(s) Summary
Update waiver entries
tests/integration/test_lists/waives.txt
Adds a skip entry for test_nvfp4_gather_grouped_gemm_act_fusion_blackwell and removes the skip entry for unittest/_torch/thop/parallel. Both entries reference issue 6966263.

Priority: ⬇️ Low

Estimated code review effort: 1 (Trivial) | ~5 minutes

Change: Bug fix

Suggested reviewers: brnguyen2

Merge Risk: ⚪ Minimal · up to ab85d

The waiver update is ready to merge with no identified blocking risk.

Pre-merge checks: Passed 5
Check name Status Explanation
Title check Passed The title clearly identifies the NVBugs issue, test-related change, flaky CuTe DSL MoE NVFP4 act-fusion test, and narrowed waive.
Description check Passed The description explains why the old directory-level waive no longer matched, identifies the replacement function-level waive, and documents relevant pre-commit coverage. The required sections and che…
Docstring Coverage Passed No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…
Linked Issues check Passed Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check Passed Check skipped because no linked issues were found for this pull request.
Full details: Description check

Explanation

The description explains why the old directory-level waive no longer matched, identifies the replacement function-level waive, and documents relevant pre-commit coverage. The required sections and checklist are present.

Full details: Docstring Coverage

Explanation

No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0 files. (1 skipped: 1 unsupported.)

✨ Finishing Touches
🧪 Generate unit tests (beta)
  • Create a new PR
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Comment @coderabbitai help to get the list of available commands.

@tensorrt-cicd

Copy link
Copy Markdown
Collaborator

PR_Github #79805 [ run ] triggered by Bot. Commit: 979f086 Link to invocation

@tensorrt-cicd

Copy link
Copy Markdown
Collaborator

PR_Github #79805 [ run ] completed with state SUCCESS. Commit: 979f086
/LLM/main/L0_MergeRequest_PR pipeline #65695 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.

CI Report

Link to invocation

…sl-moe-nvfp4

Signed-off-by: Lori Ren <lorir@nvidia.com>

# Conflicts:
#	tests/integration/test_lists/waives.txt
@lori-ren

Copy link
Copy Markdown
Contributor Author

/bot skip --comment "waiver conflict, unwaived cases have passed CI already"

1 similar comment
@lori-ren

Copy link
Copy Markdown
Contributor Author

/bot skip --comment "waiver conflict, unwaived cases have passed CI already"

@tensorrt-cicd

Copy link
Copy Markdown
Collaborator

PR_Github #79998 [ skip ] triggered by Bot. Commit: ab85d2b Link to invocation

@tensorrt-cicd

Copy link
Copy Markdown
Collaborator

PR_Github #79998 [ skip ] completed with state SUCCESS. Commit: ab85d2b
Skipping testing for commit ab85d2b

Link to invocation

@lori-ren
lori-ren merged commit 5cbcc9c into NVIDIA:main Oct 23, 2026
9 checks passed
leslie-fang25 added a commit to leslie-fang25/TensorRT-LLM that referenced this pull request Oct 23, 2026
…ct-fusion relu2 test

The kernel computes the activation and per-block amax in fp32, while the
reference rounds the GEMM output and the Relu2 output to bf16 before
fp4_quantize. Blocks whose scale factor sits near an E4M3 rounding
midpoint then round to the adjacent code, which affects 2-5% of the
blocks for Relu2, so the 95% exact-match threshold failed intermittently
(~2.6% of runs for 8192-8-1-128-relu2). Lower it to 90%.

Also remove the waiver of this test (nvbug 6966263, re-added at its
moe/kernels path by NVIDIA#20800); nvbug 6971596 reports the same relu2 SF
mismatch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: leslie-fang25 <leslief@nvidia.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants