[CI] Align extraction test with canonical auxiliary layer order

[khluu]

The H200 MIG hidden-state extraction lane fails its predictable-model layer assertion and then cannot start the following Qwen3.5 test because the first test's engine retains GPU memory (main build 87376).

#50514 intentionally canonicalizes auxiliary layer IDs by sorting and deduplicating them. The integration test supplies [5, 2, 10] but still expects the output in that input order, while the model now returns [2, 5, 10]. Keep the unsorted input and check all three extraction scenarios against canonical order, consistent with the existing auxiliary-layer unit test. Run the predictable-model test in its own process using the existing helper so engine memory is released before the next test.

Duplicate checks: open-PR searches for extraction auxiliary hidden sorted and extract_hidden_states found no equivalent ordering/process-isolation fix; existing extraction PRs address separate cache layout, model coverage, and documentation changes.

Validation:

• All applicable pre-commit hooks passed on commit; git diff --check passed.
• VLLM_TARGET_DEVICE=cpu .venv/bin/python -m pytest -q tests/v1/worker/test_eagle3_aux_hidden_states_pp.py: 2 passed, confirming the established sorted/deduplicated contract.
• H200 GPU 2 on h200-ci-1, existing main image 784cac7c424db2f2fdc879b672abad7e231f5698, with this test and current-main vllm/model_executor/models/interfaces.py mounted: /tmp/.venv/bin/python -m pytest -v tests/v1/kv_connector/extract_hidden_states_integration/test_extraction.py -k 'predictable or qwen35': 2 passed, 1 deselected in 132.69s. The venv was created with uv venv --system-site-packages /tmp/.venv.
• This is a test-only change; no model behavior changed. The exact 18 GiB MIG lane still needs CI validation: the local check used a full H200 and an existing image with the current interface overlaid, rather than a complete rebuild of current main.

AI assistance was used for investigation, implementation, and validation. Draft for human review and exact-lane CI; human line-by-line review has not been represented as completed.

CI follow-up: Buildkite 87388 was triggered for this PR head. The exact regression job(s) are enabled: (H200 MIG 18GB) Extract Hidden States Integration. At 2026-09-05T10:52:05+00:00, they were waiting for image/dependency jobs; no successful end-to-end result is claimed.

[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

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[khluu]

/ci run

[github-actions]

✅ Triggered Buildkite CI #87388 for commit 514ae2a11147.

[AndreasKaratzas]

@khluu Can we merge this? It also resolves a failure in AMD CI.

[yewentao256]

LGTM, thanks for the work!

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.