[CI] Synchronize shared offload unlink test before observing pathname

[khluu]

The V1 Core KV Offload lane failed in build 87386 because test_mp_barrier_unlinks_file_and_survives_sigkill observed the mmap pathname before its creator unlinked it. The constructor's barrier synchronizes mapping, then the creator unlinks; a peer may return and inspect the pathname while that unlink is still pending.

Synchronize the test workers again after construction, before reporting the pathname state. Keep the existing assertions for a shared inode, pathname removal, and mappings surviving until SIGKILL. Production behavior is unchanged.

Duplicate checks: searched open PRs for shared_offload_region and unlink barrier. #53073 changes creator ownership, #54619 reaps orphaned regions, and #51317 changes region reclamation. Their changes do not provide this post-constructor synchronization in the existing test.

Validation on Linux, public main commit 6865e67f0b plus this patch:

• VLLM_TARGET_DEVICE=cpu .venv/bin/python -m pytest -q tests/v1/kv_offload/cpu/test_shared_offload_region.py: 40 passed.
• Deterministic reproduction, using a temporary test-only 0.5-second delay in os.unlink: the original test failed its pathname assertion; the patched test passed with the same delay. The delay was removed afterward and is not part of this change.
• pre-commit run --files tests/v1/kv_offload/cpu/test_shared_offload_region.py: all applicable hooks passed.

CI follow-up: in build 87454 on head 08e1e2f78c, the tests/v1/kv_offload suite passed 570 tests, 1 skipped, including the previously failing test. The broader V1 Core + KV + Metrics job continues with other suites, so this is not a claim that the whole build has passed.

AI assistance was used to investigate, implement, and validate this change. Published as a draft for human review; human review is not claimed as completed. No model behavior changes or model evaluations are involved.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Team

Run ID: a968c92e-0e89-4eb9-b49b-b6a327813cdb

📥 Commits

Reviewing files that changed from the base of the PR and between 6865e67 and 08e1e2f.

📒 Files selected for processing (1)

• tests/v1/kv_offload/cpu/test_shared_offload_region.py

---

📝 Summary

Summary by CodeRabbit

• Tests
  • Updated multiprocessing barrier test synchronization to verify ordering around shared offload region setup.
  • Added coverage ensuring worker views are created only after the creator completes the required file-handling step.

Walkthrough

The multiprocessing test worker now performs a second barrier wait after SharedOffloadRegion construction and before creating its worker view.

Changes

Shared offload synchronization test

Layer / File(s)	Summary
Barrier synchronization checkpoint tests/v1/kv_offload/cpu/test_shared_offload_region.py	The worker waits at a second barrier after constructing SharedOffloadRegion and before creating its worker view.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Suggested reviewers: etelis

---

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 #87454 for commit 08e1e2f78cc0.

[pavelzak]

+1, my NVFP4 PR #52501 hit the same failure in the H200 V1 Core + KV + Metrics lane: Buildkite #87502, commit 7b47e5428.

test_mp_barrier_unlinks_file_and_survives_sigkill failed at line 939 on assert not any(r["path_exists"] for r in results). Both workers reported the same inode, and the log showed the creator unlinking the file. The KV-offload suite finished with 1 failed, 569 passed, 1 skipped.

That PR only changes NVFP4 scale handling in modelopt.py, so this is another occurrence blocking unrelated work. Thanks for addressing the race.

[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.