ci(clef): fix offline lookup and defer 27B Community cases

[yifeif-nv]

Background

Community CI stages checkpoints before launching credential-free GPU containers, but Clef's fixture still queried Hub metadata in offline mode. The subsequent L4 run got past lookup and Docker reported OOM during validation. The selected checkpoint is the complete 27B Clef release, outside the intended small-model Community scope.

Exit Criteria

Consume staged checkpoints without metadata requests. Explicitly defer the three premerge 27B cases from Community while retaining their original validation for the existing Internal/Nightly lanes. Community must still execute real GPU baseline cases and show Clef as deferred, not passed.

Implementation

The family fixture passes local_files_only according to the offline environment. Three 27B cases now declare community_gpu: false; a generic metadata check accepts only boolean values for that optional field. Model code, inputs, references, premerge flags and every original numerical assertion remain unchanged. The deployed Dev planner from #1620 supplies mandatory baseline coverage for an otherwise fully deferred selection.

Change categories

• CI or developer tooling

Validation

Commands and Results

• Matched Hub 1.33.0 / NumPy 2.4.6 offline control against base 48765e7d: 5 failed, 1 passed before the lookup fix.
• python -m pytest families/clef/tests -q -p no:cacheprovider: 20 passed, 9 existing unselected GPU/E2E cases skipped on the final source tree.
• Closed-manifest and premerge metadata guards: 2 passed. Ruff and diff checks passed; removing only the three new flags restores the original manifest, including premerge and thresholds. Production model code and original E2E assertions are unchanged.
• Earlier source head 55417d92 passed Stable CPU 37775397380: 1,535 Python tests passed, 2 skipped; 246/246 CTests passed.
• Its earlier Dev 37775427586 built native probes and passed 20 CPU plus 2 small hardware tests, then Docker reported OOM with no valid verdict for the three full-model cases. Owner and independent cleanup both confirmed deletion; this was not a parity result.

Hardware, Environment, and Revisions

Final head a9fa6f565b01fe5836e4f91b60393d5692864760, based on main 48765e7df0b7dbbe03a21d8b63f8c58db691ab3d. Local controls used a network-disabled Linux ARM64 CPU container. The historical GPU run used AWS g6.4xlarge / L4 and Dev f9343f586d04e7425c8588ee6a60b893921fee86; that older run is not final-head baseline qualification.

Not Run / Remaining Gaps

Final-head Dev 37795815020 succeeded on CI 13947f20: 1,540 Python tests passed, 2 skipped; 246/246 CTests passed; all 9 baseline GPU E2Es passed with zero selected skips. Clef was explicitly deferred, with no Clef checkpoint staging or container. The default 64 GiB profile selected AWS g6.4xlarge; both cleanup paths confirmed deletion. The 27B model's complete GPU validation is deliberately outside Community coverage; no Nightly or Internal premerge pass is claimed. This PR does not qualify Clef-Flash or modify #1608's branch.

Contributor Self-Review

• I have completed a self-review of this change.

Notes For Future Readers

Depends on the explicit deferral/baseline policy deployed to Dev in #1620. Preserve the offline lookup behavior when integrating #1608's per-checkpoint fixture. A successful Community status validates its runnable GPU baseline, not the deferred Clef model.

Risk level

• Medium

Coverage routing changes are explicit. Original model criteria and premerge declarations remain intact, and the generic planner rejects an empty green result by requiring real baseline cases.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

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