Skip to content

ci(hipblaslt): move ASAN tests to the gfx90a OSS CI pool - #12813

Draft
newling wants to merge 2 commits into
developfrom
users/newling/hipblaslt-asan-test-runner
Draft

newling wants to merge 2 commits into
developfrom
users/newling/hipblaslt-asan-test-runner

Conversation

@newling

@newling newling commented Sep 30, 2026 •

Copy link
Copy Markdown
Contributor

hipBLASLt's gfx90a AddressSanitizer (ASAN) smoke job can hang during library initialization, before any tests start. The symptom is tracked in #12727, and ROCm/TheRock-Infra#978 tracks the investigation of the linux-gfx90a-gpu-rocm runner pool.

This PR changes the test runner to linux-gfx90a-1gpu-ossci-rocm. It changes one runner label and can land independently of #12716 and #12720. The 20-minute test timeout and runner diagnostics from the already-merged #12728 remain in place.

In the controlled runner comparison, the same artifact, container and ASAN settings hung before the version banner on the original pool and completed the smoke suite on the selected pool. Two debugger captures showed the stalled process waiting for GPU code-cache invalidation during the first hipMemset in handle creation. The specific host/runtime cause remains unresolved. The original pool also has intermittent passes, but another startup timeout on October 5 shows that the symptom persists.

The full ASAN workflow passed both before the rebase and after the rebase. The latter completed all 948 smoke tests. Reusing that artifact, a diagnostic pilot and the first qualification trial also passed the complete smoke suite and normally linked HIP and hipBLASLt startup probes on two distinct physical GPU sets. Their original-pool counterparts remain queued, so the planned ten-pair comparison has no completed pairs yet.

The latest commit, c65f79d4d37, adds only the inline workaround explanation and issue link. Root pre-commit and actionlint pass; its fresh CI run is queued.

This is a runner-pool workaround. Repeated runs are still needed to establish reliability, and #12727 and ROCm/TheRock-Infra#978 remain open for investigation of the underlying cause.

@github-actions github-actions Bot added github actions project: none Does not target any component labels Sep 30, 2026
@therock-pr-bot

therock-pr-bot Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

✅ All Checks Passed — Ready for Review

Check Status Details
📝 PR Description ✅ Pass —
⛔ Forbidden Files ✅ Pass —
🧪 Unit Test ✅ Pass PR does not contain code files — Unit Test auto-passed
🔎 pre-commit ✅ Pass —
🚫 Draft PR 🔜 To Be Enabled —
🚩 Feature Flag 🔜 To Be Enabled —
📊 Code Coverage 🔜 To Be Enabled —
🤖 therock-pr-bot ✅ Pass —

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

@therock-pr-bot

therock-pr-bot Bot commented Sep 30, 2026 •

Copy link
Copy Markdown

🎉 All checks passed! This PR is ready for review.

@newling
newling marked this pull request as ready for review September 30, 2026 23:03
@newling
newling requested a review from a team as a code owner September 30, 2026 23:03
@newling
newling force-pushed the users/newling/hipblaslt-asan-test-runner branch from 55deb15 to d820382 Compare October 5, 2026 17:49
@newling
newling force-pushed the users/newling/hipblaslt-asan-test-runner branch from c65f79d to 337340b Compare October 6, 2026 22:01
sa-faizal added a commit that referenced this pull request Oct 8, 2026
ISSUE ID: #12727

## Motivation
The hipBLASLt per-PR ASAN lane has been hanging under full ASAN with
`HSA_XNACK=1`. This switches it to host-only ASAN with XNACK off,
matching the coverage already described in `TESTING.md`.

GPU correctness tests still run, and ASAN still checks CPU memory
accesses. The trade-off is that no gfx90a device-ASAN runtime tests
remain in CI. Before restoring that coverage, we need to review runner
capacity and full-ASAN stability with DevOps, potentially for a
dedicated nightly pool. The cause of the hangs remains under
investigation.

[#12728](#12728) added
shorter timeouts and diagnostics.
[#12813](#12813) offers an
alternative runner-pool workaround. The investigation is tracked in
[TheRock-Infra #978](ROCm/TheRock-Infra#978).

## Technical Details
- Use `HOST_ASAN` artifacts and `HSA_XNACK=0` in the focused lane.
- Stop requiring `xnack+` targets for host-only ASAN; preserve existing
full-ASAN behavior.
- Update `TESTING.md` and remove the stale gfx942 `ci:asan` row.

## Test Plan
Run the existing gfx90a build and smoke tests, plus whitespace and
applicable pre-commit checks.

## Test Result
[Build and all 950 smoke tests
passed](https://github.com/ROCm/rocm-libraries/actions/runs/37402899925)
on `c92deaa`. The documentation follow-up passes whitespace checks;
pre-commit completed with no applicable hooks. CI for the latest commit
must pass before merge.

## Submission Checklist
- [x] Reviewed the [contributing
guidelines](https://github.com/ROCm/TheRock/blob/main/GOVERNANCE.md#pull-requests).

## Risk level
Medium (3/5): retains host-ASAN and GPU correctness checks, but removes
device-ASAN coverage from the gfx90a lane. A passing run does not prove
the hangs are fixed.

---------

Co-authored-by: Sahil Faizal <197828758+sa-faizal@users.noreply.github.com>
Co-authored-by: Sahil Faizal <safaizal@amd.com>
@newling

newling commented Oct 8, 2026

Copy link
Copy Markdown
Contributor Author

The ASAN CI job is now semi-fixed (failing component is removed from it) so this PR is less relevant.

@newling newling closed this Oct 8, 2026
@newling newling reopened this Oct 9, 2026
@newling
newling marked this pull request as draft October 9, 2026 23:55
@newling
newling force-pushed the users/newling/hipblaslt-asan-test-runner branch 2 times, most recently from cc0a8c2 to ce9db58 Compare October 13, 2026 03:46
@newling
newling force-pushed the users/newling/hipblaslt-asan-test-runner branch from ce9db58 to 49a339f Compare October 26, 2026 15:25

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant