[dileepr1]

Motivation

GitHub issue: #7202

The libraries_baseline_ready job added in #8069 exists for one purpose: to let bump_automation.py decide whether a completed run's reused stages are good enough to hand out as a baseline_run_id. #8069 assessed it as inert:

The new libraries_baseline_ready job is additive: nothing else in Multi-Arch CI depends on it, so a bug in it can make its own status wrong but cannot block or skip any other job in the run.

That does not hold in practice. The job lives inside the reusable workflows multi_arch_build_portable_linux.yml / multi_arch_build_windows.yml, so when it fails, the calling build_multi_arch_stages job is marked failed. test_artifacts_per_family is guarded by if: ${{ !failure() && !cancelled() }} and therefore never starts — the test matrix does not even expand.

This bites any caller that narrows the build graph with build_stages. The stages it deliberately excluded report skipped, the gate treats skipped as not-success and fails, and the run goes red having run no tests. Observed on the rocm-systems host-asan presubmit in https://github.com/ROCm/rocm-systems/actions/runs/34608811109, where both requested stages built cleanly:

Job	Result
compiler-runtime / Stage - Compiler Runtime	success
runtime-tests / Stage - Runtime Tests	success
the other 8 library stages	skipped (excluded via build_stages)
Libraries Baseline Ready	failure
Test ${{ matrix.family_info.amdgpu_family }}	skipped — matrix never expanded

Worth noting the gate fired in a context nothing will ever read: get_baseline_run_id_from_merged_pr() only scans TheRock's own Multi-Arch CI runs after a merge, never an external repo's run.

Technical Details

Dropped if: always() and the conditional failing step. GitHub's default needs: semantics already express exactly the intent — a job whose needs: did not all succeed is skipped — so the gate now runs only when every reused stage succeeded, and is skipped otherwise.

The rejection signal is unchanged from the bump automation's point of view. _baseline_gate_jobs_succeeded() keys on conclusion == "success", and skipped jobs are still returned by the Actions jobs API with conclusion: "skipped", so a skipped gate rejects the baseline exactly as a failed one did. All three cases still land correctly:

• every reused stage green → gate succeeds → baseline usable
• any reused stage failed, cancelled or skipped → gate skipped → baseline rejected
• run no longer goes red, so downstream tests are unaffected

The remaining step now just echoes toJSON(needs) for diagnosis when the gate does run.

Test Plan

• python -m pytest build_tools/github_actions/tests/bump_automation_test.py -q
• pre-commit run --files .github/workflows/multi_arch_build_portable_linux.yml .github/workflows/multi_arch_build_windows.yml build_tools/github_actions/tests/bump_automation_test.py

Test Result

63 passed. Targeted pre-commit passed, including actionlint on both workflow files.

Risk Assessment

Risk 2. CI wiring on the libraries/systems bump path. The change strictly reduces what the gate can break: it can no longer fail, so it can no longer propagate failure to a calling job. The baseline-rejection behavior the gate was added for is preserved, and is now covered by a test asserting a skipped gate job is rejected.

Related

• Introduced the gate: fix(ci): fix TheRock submodule bump pin-PR automation for rocm-libraries and rocm-systems #8069 (work tracking Libraries TheRock pin bump PRs omit Multi-Arch workflow uses: pins #8068)
• Unblocks: [issue] [asan] Support runtime-only host-ASan presubmit builds and tests #7202 host-asan presubmit, which relies on build_stages narrowing

Device / Architecture Coverage

No device code. CI workflow and unit-test change only.

Testing Summary

Added test_false_when_the_gate_job_was_skipped, covering the path that is now the primary rejection route: a gate job skipped because a reused stage did not succeed must not be handed out as a baseline. Existing coverage for succeeded / failed / missing / multi-platform gate jobs is unchanged and still passes.

Testing Checklist

• Bump automation unit tests - python -m pytest build_tools/github_actions/tests/bump_automation_test.py -q - Status: Passed (63)
• Targeted pre-commit incl. actionlint - Status: Passed
• PR CI - GitHub PR checks - Status: Pending

Flags / Guardrails

None.

Adjacent Tests Considered

GetBaselineRunIdFromMergedPrTest exercises the lookup end to end against mocked job lists and needed no change, since it already drives _baseline_gate_jobs_succeeded through non-success conclusions.

Risk Acceptance / Waivers

None.

Technical Changes

• Remove if: always() and the conditional exit 1 step from libraries_baseline_ready in both the Linux and Windows multi-arch build workflows, so the job is skipped rather than failed when a reused stage did not succeed.
• Add a unit test asserting a skipped gate job is rejected as a baseline.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/TheRock/blob/main/CONTRIBUTING.md.

[therock-pr-bot]

✅ 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]

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

[tony-davis]

Hit the same bug from the other direction today: ROCm/rocm-libraries#11952 (a bump-bot PR using prebuilt_stages/baseline_run_id reuse) went red on both the Linux and Windows Libraries Baseline Ready jobs for exactly this reason, on the same commit that introduced the gate (#8069). Good catch that it also breaks build_stages-narrowed runs like the host-asan presubmit; I hadn't considered that case.

I like this fix better than the per-stage contains(inputs.prebuilt_stages, ...) check I was about to write. Letting the default needs: semantics do the work is simpler and doesn't need updating every time the reused-stage list changes.

Checked out the branch and ran it locally:

• pytest build_tools/github_actions/tests/bump_automation_test.py -q → 63 passed
• pre-commit run --files .github/workflows/multi_arch_build_portable_linux.yml .github/workflows/multi_arch_build_windows.yml build_tools/github_actions/tests/bump_automation_test.py → all green, including actionlint

LGTM, approving.