[Bugfix] Fix faulty triton importing logic when using Ray for DP #19734
Conversation
Signed-off-by: mgoin <mgoin64@gmail.com>
|
👋 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 @mgoin, 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 that prevented Triton from being correctly initialized and utilized when vLLM is run in a data parallel setup using Ray. The fix introduces a more robust check for Triton driver availability, specifically accounting for how distributed environments like Ray temporarily modify CUDA_VISIBLE_DEVICES during actor initialization, ensuring Triton is properly enabled.
Highlights
- Bugfix: Resolves a critical bug where Triton was incorrectly disabled during vLLM's initialization when running in a data parallel setup with Ray. This was caused by Ray temporarily setting
CUDA_VISIBLE_DEVICESto an empty string, which the previous Triton import logic misinterpreted as a lack of active drivers. - Triton Import Logic Enhancement: The logic for checking Triton driver availability has been made more robust. It now specifically accounts for distributed environments (like Ray) where
CUDA_VISIBLE_DEVICESmight be temporarily empty during actor initialization, allowing Triton to correctly initialize in such scenarios. - Environment Variable Handling: Introduces a check for the
CUDA_VISIBLE_DEVICESenvironment variable to identify if the process is operating within a distributed setup where temporary empty values are expected, thus preventing premature Triton disabling.
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 where Triton was being incorrectly disabled in distributed environments due to the timing of the CUDA_VISIBLE_DEVICES environment variable being set by Ray. The changes introduce a check for this specific scenario and adjust the driver check accordingly. I've suggested a minor readability improvement and a more descriptive debug message.
| is_distributed_env = (cuda_visible_devices is not None | ||
| and len(cuda_visible_devices.strip()) == 0) |
There was a problem hiding this comment.
Would something like is_ray_env be better here? Since presumably this will be False in non-ray distributed cases?
There was a problem hiding this comment.
btw, an alternative is to check Ray directly:
Line 2586 in 5a1c2e1
…m-project#19734) Signed-off-by: mgoin <mgoin64@gmail.com>
…m-project#19734) Signed-off-by: mgoin <mgoin64@gmail.com> Signed-off-by: minpeter <kali2005611@gmail.com>
…m-project#19734) Signed-off-by: mgoin <mgoin64@gmail.com> Signed-off-by: Yang Wang <elainewy@meta.com>
…m-project#19734) Signed-off-by: mgoin <mgoin64@gmail.com>
Purpose
FIX #19731
In the
DPEngineCoreActor, Ray setsCUDA_VISIBLE_DEVICESto an empty string, and then the code deletes it to properly initialize data parallel groups.vllm/vllm/v1/engine/core.py
Lines 922 to 925 in aed8468
However, this happens after the Triton driver check has already run during the module import.
The key issue is the timing:
vllm.triton_utils.importingmodule is imported firstCUDA_VISIBLE_DEVICESis set to empty string by RayDPEngineCoreActor.__init__()method deletes the problematicCUDA_VISIBLE_DEVICESenvironment variableTest Plan
Manually verify
TP_SIZE=1 DP_SIZE=2 pytest -s -v "v1/test_async_llm_dp.py::test_load[ray-RequestOutputKind.DELTA]"works now.Test Result
Passed!