Repository navigation
[https://nvbugs/7078630][test] Unwaive FP4 MLA helix-scatter unit tests fixed by #21313 - #22421
reasonsolo wants to merge 1 commit into
Conversation
…ts fixed by NVIDIA#21313 Signed-off-by: reasonsolo <1432185+reasonsolo@users.noreply.github.com>
|
No actionable comments were generated in the recent review. WalkthroughThe skip list no longer includes the active and inactive FP4 MLA Helix scatter test cases. ChangesTest skip list
Estimated code review effort: 1 (Trivial) | ~3 minutes Change: Other Suggested reviewers: Merge Risk:
|
sunnyqgg
left a comment
There was a problem hiding this comment.
0 blockers, 0 possible blockers, 0 required, 1 nice to have. The unwaive looks right. The fix (#21313, d7f2a84) is on this branch's base. The waived failure was deterministic: in post-merge 3066–3069 both cases failed at the old whole-buffer prequantized_q assertion (torch.Size([128, 320]) != torch.Size([4, 320]), from the packed-Q padding), which #21313 corrected in the test. The remaining assertions match the code at the base: Q batch capacity 1, RoPE at the Helix global position, and inactive ranks return before any cache store. Nothing in the FP4 MLA path or the test changed between the cited run's commit 1e658cb and the base. The only ask is a CI run of the post-merge GR100 stage on this PR (inline).
| @@ -169,8 +169,6 @@ test_doc.py::test_http_url_validity SKIP (https://nvbugs/7066122) | |||
| test_e2e.py::test_ptp_quickstart_bert[VANILLA-BertForSequenceClassification-bert/bert-base-uncased-yelp-polarity] SKIP (bug pending, tracked in PR 17414) | |||
| unittest/_torch/attention/multi_gpu --ignore=unittest/_torch/attention/multi_gpu/test_helix_zero_kv.py SKIP (https://nvbugs/7044056) | |||
| unittest/_torch/attention/sparse/csa2/test_module.py::test_fused_native_index_q_and_graph SKIP (https://nvbugs/7039506) | |||
There was a problem hiding this comment.
[Nice to have] Please link a CI run for the two cases unwaived just below. test_fp4_mla.py runs only on GR100-PyTorch-Post-Merge-1 (l0_gr100.yml, stage: post_merge; every other list skips it via _is_cutedsl_unavailable()), so the default /bot run won't execute these cases. Every earlier CI run (post-merge 3066–3069) stopped at the old whole-buffer prequantized_q assertion, so the later assertions have never run in CI. Suggest /bot run --stage-list "GR100-PyTorch-Post-Merge-1" (an explicitly named post-merge stage doesn't need the ci: post-merge approved label), then check that both [active] and [inactive] show PASSED, not SKIPPED, in the inner test_fp4_mla.py results. Go by those inner results: in post-merge 3076 the stage also had unrelated MiniMax-M3 and test_moe_backend failures.
|
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. |
|
/bot run --stage-list "GR100-PyTorch-Post-Merge-1" |
|
PR_Github #85347 [ run ] triggered by Bot. Commit: |
What
Removes the two stale SKIP entries from
tests/integration/test_lists/waives.txt:unittest/_torch/attention/test_fp4_mla.py::test_fp4_mla_helix_scatter_uses_global_rope_and_local_cache_position[active]unittest/_torch/attention/test_fp4_mla.py::test_fp4_mla_helix_scatter_uses_global_rope_and_local_cache_position[inactive]Both were waived for https://nvbugs/7078630. Unwaive only — no production code changes.
Why
The helix-scatter fix landed in
mainvia #21313 (d7f2a84c), and post-merge verification shows both parametrizations pass:1e658cb1(containsd7f2a84c)test_unittests.pyadapterThat verification covers these two cases only, not the other changes in #21313 and not performance.
Scope and safety
Unwaive only.
tests/unittest/_torch/attention/test_fp4_mla.pykeeps its hardware guard (_is_cutedsl_unavailable()skips unless Rubin SM107 + CuTeDSL runtime), and the file is already part oftests/integration/test_lists/test-db/l0_gr100.yml, so the cases simply leave the waive list.Verification of this change
python3 scripts/check_test_list.py --validate --check-duplicate-waives→OK: 2228 unique test entries validatedgit diffis a 2-line deletion inwaives.txt; no othernvbugs/7078630reference remains in the treeWaiver-tracking follow-up for nvbugs/7078630.
Dev Engineer Review
SKIPentries fromtests/integration/test_lists/waives.txt. The entries cover the active and inactive parametrizations oftest_fp4_mla_helix_scatter_uses_global_rope_and_local_cache_position.#21313being inmain. The PR text states this, but this review did not verify it.check_test_list.py --validate --check-duplicate-waivesand the pre-commit checks passed. These results are author-reported and were not re-run for this review.QA Engineer Review
Test-list files modified:
tests/integration/test_lists/waives.txt: removed theSKIPentries for the active and inactive cases oftest_fp4_mla_helix_scatter_uses_global_rope_and_local_cache_position.Behavior covered:
tests/integration/test_lists/test-db/l0_gr100.ymlalready lists the test file. This review did not verify either claim.Coverage verdict:
sufficientfor the two unwaived cases, based on the author's report of a passing run on one GR100 GPU. The result does not cover other changes in#21313or performance.Per-File QA Perspective
tests/integration/test_lists/waives.txt: Removing these two skips makes the FP4 MLA helix-scatter cases run in CI. QA should confirm that both cases pass on the GR100 CI pool.