Repository navigation
test(hipblaslt): quarantine the gfx950 alpha = 0, beta = 1 complex batch edge cases (ROCM-33823) - #14867
Merged
tony-davis merged 2 commits intoNov 2, 2026
Conversation
…tch edge cases (ROCM-33823) Since #14743 fills D with a NaN sentinel before each verified launch, the gfx950 kernels that leave some D elements unwritten in matmul_batch_complex_general_batch_edge fail every time instead of only when leftover memory happens to be NaN (#14282). That blocks every TheRock rocm-libraries bump since ROCm/TheRock#9651. Quarantine the alpha = 0, beta = 1 cases on gfx950 until ROCM-33823 is fixed. Co-authored-by: Cursor <cursoragent@cursor.com>
✅ All Checks Passed — Ready for Review
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
|
🎉 All checks passed! This PR is ready for review. |
Contributor
There was a problem hiding this comment.
🟢 Approval recommended
No active review issues were identified for the gfx950-only quarantine of the tracked complex batched edge-case failures.
No open issues 🎉
What changed in this PR
Quarantines the known gfx950 complex batched edge-case failures tracked by ROCM-33823, which became reproducible after verification began resetting D with a NaN sentinel.
Changes:
- Adds a gfx950-only known-bug entry for
matmul_batch_complex_general_batch_edgewith alpha = 0 and beta = 1, both with zero imaginary components. - Documents the failure, removal condition, and November 16 review deadline.
| File | Description |
|---|---|
projects/hipblaslt/clients/tests/data/known_bugs.yaml |
Adds the scoped quarantine and tracking comments. |
Review effort: Balanced
newling
approved these changes
Nov 2, 2026
…mplex-alpha0-quarantine
tony-davis
enabled auto-merge (squash)
November 2, 2026 17:11
Copilot started work to fix failing Actions workflows on behalf of
tony-davis
November 2, 2026 17:11
tony-davis
added
ci:skip
tony-davis
deleted the
users/todavis_amdeng/ROCM-33823-gfx950-complex-alpha0-quarantine
branch
November 2, 2026 20:04
Contributor
TheRock Submodule Bump ActivityNewest first
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Jira ID: ROCM-33823
Motivation
Every TheRock rocm-libraries bump since ROCm/TheRock#9651 fails gfx950 hipBLASLt shard 4 on
matmul_batch_complex_general_batch_edge(cf64 TN 1x1024x512, alpha = 0, beta = 1, -nan in D). This was the intermittent #14282. #14743 now fills D with a NaN sentinel before every verified launch, so a kernel that leaves D elements unwritten fails every time. That's the fix working as intended, and it's exposing a real kernel bug tracked in ROCM-33823.Technical Details
Adds one known_bugs.yaml entry that quarantines the alpha = 0, beta = 1 cases of that test on gfx950 only. I scoped it to the alpha/beta pair rather than the single failing case because the bug has moved between shapes and types (cf64 NN and TN, cf32 NN and NC). All other alpha/beta combinations and all other arches still run.
Test Plan
Ran hipblaslt_gentest.py over matmul_gtest.yaml and confirmed the entry marks only the alpha = 0, beta = 1 cases of this test for gfx950. CI.
Test Result
TBD (CI)
Submission Checklist