[bugfix] support safe_dispatch env to use safe branch in pd_router #9502
Conversation
There was a problem hiding this comment.
Summary of Changes
Hello @Bruce-x-1997, 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 introduces an optional "safe dispatch" mechanism within the pd_router to enhance stability, particularly in online environments where sgl-router has experienced issues. By setting a new environment variable, users can opt into a more robust handling of prefill and decode responses, aligning the behavior with the more stable minilb approach.
Highlights
- Environment Variable Control: A new environment variable, SGL_ROUTER_SAFE_DISPATCH, is introduced to allow users to explicitly enable a safer dispatch path.
- Conditional Safe Dispatch: The system now conditionally uses a "safe branch" for handling prefill and decode requests when SGL_ROUTER_SAFE_DISPATCH is true, or when context.return_logprob is enabled.
- Improved Concurrency Handling: The "safe branch" utilizes tokio::join to concurrently await prefill and decode responses, mirroring the stable behavior observed in minilb.
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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.
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
This pull request introduces a SGL_ROUTER_SAFE_DISPATCH environment variable to enable a safer dispatch mechanism in the PD router, which helps with stability issues observed in production. The implementation correctly uses this flag to switch to a code path that waits for both prefill and decode responses.
My main feedback is to refactor the environment variable handling. Currently, it's read on every request, which is inefficient. I've suggested moving this to the router's initialization to improve performance. I also suggested a more robust way to parse the boolean value from the environment variable for better usability.
Motivation
I met stability issue at online case when using sgl-router, and when we use minilb, it works well. After reading the code of minilb and sgl-router, I set a env var to get running-well online.
It could not be found when bench_serving, but when online, I could see many mooncake transfer fail because of abortReq, but there is no abortreq in requests.Then after communicating with community, it might because of a bug about prefill response management, then http session at prefill side is broken so it trigger abortReq in scheduler.
Modifications
I just provide a SGL_ROUTER_SAFE_DISPATCH env var, and when it's true, just run safe branch, which use tokio.join to wait prefill & decode response(reuse the return_logprobs code), the same as minilb.
From community, I know http path would be deprecated after soon, so just work around to help guys who meet the problem like me.
Checklist