[CI/Build][Bugfix] Fix deadlock on v1 engine test CI

[Isotr0py]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

• This PR fixes the deadlock in tests/v1/engine/test_async_llm.py, which caused the v1 test timeout on main.

Test Plan

pytest -v -s v1/engine

Test Result

• The test should pass now.

(Optional) Documentation Update

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @Isotr0py, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request addresses a critical CI issue by resolving a deadlock that was occurring during the execution of a specific V1 engine test. The fix ensures the test completes successfully, preventing timeouts and contributing to a more stable continuous integration pipeline.

Highlights

• Bugfix: Fixed a deadlock issue occurring in the test_check_health function within tests/v1/engine/test_async_llm.py. This deadlock was causing CI timeouts on the main branch.
• CI Stability: Improved the stability of the CI pipeline by resolving a specific test failure related to engine initialization and threading.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

The pull request aims to fix a deadlock in the test_check_health CI test for the v1 engine. The change involves wrapping the AsyncLLM engine initialization with set_default_torch_num_threads(1).

This modification is consistent with a pattern observed in multiple other tests within the test_async_llm.py file (e.g., test_load, test_abort, test_finished_flag). Using set_default_torch_num_threads(1) is a common strategy to mitigate threading-related issues, such as deadlocks or race conditions, in test environments by ensuring PyTorch operations are single-threaded during sensitive phases like engine setup.

The change is minimal, targeted, and appears to be a direct and appropriate fix for the described CI deadlock. No issues of medium, high, or critical severity were identified in the applied diff. The code adheres to general Python best practices.

[njhill]

Thanks for fixing these @Isotr0py.

This with set_default_torch_num_threads(1) thing seems like a hack (I just caught up on the original changes in #19316). Do we understand why it's needed? It seems fragile/incorrect to require all of the tests to have this (and if the tests require it presumably users would too?)

[Isotr0py]

TBH, I haven't taken deeper investigation in #19316 after determined the CI green.

Do we understand why it's needed?

According to traceback in #19316 (comment), I guess the root issue should be the introduction of auto dtype casting in processor, which will call blocking cpu ops like x.to(dtype) too early during worker initialization in compute_encoder_budget.

So I diasbled OpenMP through with set_default_torch_num_threads(1) to avoid deadlock when initialize engine in CI. But I need to double check this guess.

if the tests require it presumably users would too?

If the above guess is true, I think we can only disable OpenMP temporarily through with set_default_torch_num_threads(1) when calling compute_encoder_budget during worker initialization. But I'm not sure if it will slow down the engine stratup... WDYT?

[njhill]

Thanks @Isotr0py. So this was likely introduced by #18862?

Yes it seems we should fix the root cause or at least address it internally along the lines of what yous suggest.

[Isotr0py]

So this was likely introduced by #18862?

Yes, in fact, this is the same issue mentioned in #18862 (comment) firstly, but the dtype casting function introduced in final version wouldn't take effect exactly, so we mistook that the deadlock issue has been solved at that time when the CI turn green.

And because that PR's casting function took no effect, I created #19316 to fix the dtype casting, and this unsolved deadlock issue also occurred once again at that time.