[BugFix] Fix multi-node offline data parallel #19937
Conversation
The offline data parallel example examples/data_parallel.py was broken because there's no outside synchronization tied to the engines running state, so other processes would fail once the first one finishes its requests and exits. This PR adds logic for the engines running state to be propagated to the front-end process in the offline case and taken into account by the LLMEngine.has_unfinished_requests() method. Signed-off-by: Nick Hill <[email protected]>
njhill
requested review from
WoosukKwon,
robertgshaw2-redhat,
ywang96 and
comaniac
as code owners
|
👋 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 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 🚀 |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @njhill, 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 bug in the multi-node offline data parallel execution, where processes would prematurely terminate due to a lack of proper synchronization of the engine's running state. The changes ensure that the LLMEngine accurately reflects whether the underlying engines are still active, preventing early exits and improving the robustness of distributed offline inference.
Highlights
- Bug Fix for Offline Data Parallel: Implemented a mechanism to correctly propagate the
engines_runningstate from the core engine to the frontendLLMEnginein multi-node offline data parallel setups. This prevents premature termination of processes when one rank finishes its requests before others. - Enhanced Engine State Management: Introduced
dp_engines_runningmethods and state tracking withinEngineCoreClientandEngineCoreClientImplto accurately reflect the activity of data parallel engines, ensuringLLMEngine.has_unfinished_requests()provides a reliable status. - Improved Test Coverage: Added a new multi-node test case for the
examples/offline_inference/data_parallel.pyexample to the CI pipeline, validating the fix and ensuring future regressions are caught. - Minor UI Improvement: Added a refresh call for the tqdm progress bar in
_run_engineto ensure it updates correctly upon completion of all requests.
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
-
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. ↩
There was a problem hiding this comment.
Code Review
The pull request addresses a bug in the offline data parallel example where processes would fail due to lack of synchronization. The changes propagate the engine's running state to the front-end process and incorporate it into the LLMEngine.has_unfinished_requests() method. The review comments suggest improvements to code readability through additional comments and docstrings.
| client_index = -1 if self.has_coordinator else 0 | ||
| self.output_queue.put_nowait( | ||
| (-1, | ||
| (client_index, | ||
| EngineCoreOutputs(wave_complete=self.current_wave))) |
There was a problem hiding this comment.
Consider adding a comment explaining why client_index is set to -1 when has_coordinator is true, and 0 otherwise. This will improve code readability.
| client_index = -1 if self.has_coordinator else 0 | |
| self.output_queue.put_nowait( | |
| (-1, | |
| (client_index, | |
| EngineCoreOutputs(wave_complete=self.current_wave))) | |
| client_index = -1 if self.has_coordinator else 0 | |
| # In the coordinator case, dp rank 0 sends updates to the | |
| # coordinator. Otherwise (offline spmd case), each rank | |
| # sends the update to its colocated front-end process. |
Signed-off-by: Nick Hill <[email protected]>
Signed-off-by: Nick Hill <[email protected]>
The offline data parallel example
examples/data_parallel.pywas broken because there's no outside synchronization tied to the engines running state, so other processes would fail once the first one finishes its requests and exits.This PR adds logic for the engines running state to be propagated to the front-end process in the offline case and taken into account by the
LLMEngine.has_unfinished_requests()method.Fixes #17241
Fixes #19937