[WIP] Introduce flashinfer allreduce in MoE TP #7946
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
There was a problem hiding this comment.
Summary of Changes
Hello @yuan-luo, 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 a performance optimization for Mixture-of-Experts (MoE) layers in SGLang by integrating FlashInfer's custom allreduce operation. This change aims to replace the default PyTorch distributed all-reduce for specific scenarios, particularly for smaller token counts, to potentially achieve better communication efficiency and throughput in distributed MoE models.
Highlights
- FlashInfer Allreduce Integration: A new module (
python/sglang/srt/layers/flashinfer_comm.py) has been introduced to encapsulate FlashInfer's custom allreduce operation, including necessary workspace management for inter-process communication (IPC). - Conditional MoE Optimization: The Fused Mixture-of-Experts (MoE) layer (
python/sglang/srt/layers/moe/fused_moe_triton/layer.py) is updated to conditionally utilize this new FlashInfer allreduce forfinal_hidden_states. This optimization is applied when FlashInfer is available, the new feature flag is enabled, and the input token count is 128 or less, potentially improving performance over standardtorch.distributed.all_reduce. - New Configuration Option: A new server argument
--enable-flashinfer-allreducehas been added topython/sglang/srt/server_args.py, allowing users to enable or disable this specific FlashInfer allreduce optimization for MoE layers.
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
This pull request introduces FlashInfer's custom all-reduce kernel to replace torch.distributed.all_reduce for MoE layers, aiming for performance improvements. The changes include a new file flashinfer_comm.py with the wrapper implementation, and modifications to fused_moe_triton/layer.py to conditionally use this new kernel. The server arguments are also updated to control this feature.
My review has identified several critical issues in the new flashinfer_comm.py file that need to be addressed. These include function signature mismatches, incorrect function calls that would lead to runtime errors, and the use of random dummy data in the core computation which would produce incorrect results. There is also a potential resource leak in the workspace manager. I have provided specific suggestions to fix these issues.
|
Please note this PR needs to use B200, otherwise will encounter this error. |
c2ec629 to
8fd70ad
Compare
54b40de to
15388b1
Compare
|
Encounter some problem in B200: |
kaixih
reviewed
Jul 24, 2025
5 tasks
15388b1 to
0bc6068
Compare
0bc6068 to
8d1bb24
Compare
8d1bb24 to
7cabb5d
Compare
7cabb5d to
b1d634f
Compare
|
Encounter some issues on H20. Flashinfer can't find cuda_fp4.h file. |
|
I'll verify it on B200. |
b1d634f to
50b914e
Compare
|
After install flashinfer_python-0.2.14.post1 in H20, it move forward: Hopper passed cuda graph, but the result is incorrect. Inside the custom_all_reduce API, there's a flag_value, it needs to incremented by 1 for each AR. Otherwise the result is incorrect. |
|
Check the flashinfer code, in trt-llm wrapper, it use the flag in this way, So I changed the next_flag() in this way, but still incorrect. |
|
Adding related PR：flashinfer-ai/flashinfer#1096 |
|
I tested that #6619 has better performance on custom_allreduce than trtllm-custom-allreduce used in flashinfer. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
Th
2D79
is suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Motivation
Introduce flashinfer custom allreduce into SGLang Fused MoE, replacing torch distributed all_reduce.
Modifications
Checklist