[bugfix] Fix flashinfer cutlass EP moe after MoE refactor #8630

trevor-m · 2025-07-31T19:03:07Z

Motivation

Accuracy went to 0 for flashinfer cutlass MoE after recent MoE refactor and improvements.

Modifications

Fix issue from [4/N] MoE Refactor: Unified Triton Kernel for FusedMoE and EPMoE #8515 - cutlass moe does not need to map topkid ids using expert map gpu.
Fix issue from [Feature] Hybrid EP and TP #8590 - hybrid EP/TP not yet supported for this path

Accuracy Test

python3 -m sglang.launch_server --model-path nvidia/DeepSeek-R1-0528-FP4 --trust-remote-code --quantization modelopt_fp4 --tp 8 --enable-flashinfer-cutlass-moe --enable-ep-moe --ep-size 8
python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1319 --parallel 1319 --port=30000
Accuracy: 0.961
Invalid: 0.000
Latency: 23.725 s
Output throughput: 6143.771 token/s

Benchmark & Profiling

Checklist

Format your code according to the Code Formatting with Pre-Commit.
Add unit tests as outlined in the Running Unit Tests.
Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

gemini-code-assist

Summary of Changes

Hello @trevor-m, 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 provides crucial bug fixes for the FlashInfer Cutlass MoE implementation, which was experiencing zero accuracy due to recent refactoring. The changes primarily involve refining the expert mapping logic for this specific MoE type and introducing a compatibility adjustment for expert and tensor parallelism configurations to ensure correct operation.

Highlights

FlashInfer Cutlass MoE Fix: I've addressed a critical bug where FlashInfer Cutlass MoE was causing accuracy to drop to zero after a recent MoE refactor. This involved correcting the handling of expert_map_gpu for this specific MoE path, as it was found not to require mapping top-k IDs using the GPU expert map.
Expert Parallelism (EP) and Tensor Parallelism (TP) Compatibility: I've implemented a temporary measure to handle the current lack of support for hybrid EP/TP configurations with FlashInfer Cutlass MoE. When both FlashInfer Cutlass MoE and EP MoE are enabled, the expert parallel size (ep_size) is now automatically adjusted to match the tensor parallel size (tp_size), with a warning logged to indicate this adjustment.

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 or fill out our survey 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.

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

This pull request introduces a bugfix for the flashinfer cutlass MoE implementation, which was causing accuracy issues after a recent refactoring. The changes are well-targeted and address two specific issues:

In python/sglang/srt/layers/moe/fused_moe_triton/layer.py, the code now correctly avoids mapping topkid IDs on the GPU when using the flashinfer cutlass MoE path, which is an unnecessary and incorrect operation for this specific kernel.
In python/sglang/srt/server_args.py, a check is added to enforce that the expert parallelism size (ep_size) is equal to the tensor parallelism size (tp_size) when flashinfer cutlass MoE is enabled. This prevents an unsupported hybrid EP/TP configuration.

The fixes appear correct and are consistent with the existing codebase. I have one suggestion to improve the clarity of a warning message for the user.

gemini-code-assist · 2025-07-31T19:04:45Z

python/sglang/srt/server_args.py

+                self.ep_size = self.tp_size
+                logger.warning(
+                    f"Flashinfer cutlass MoE and EP MoE are enabled. The expert parallel size is adjusted to be the same as the tensor parallel size[{self.tp_size}]."
+                )


The warning message informs the user about the automatic adjustment. It could be more informative by explaining why the adjustment is necessary. The reason, as mentioned in the PR description, is that hybrid EP/TP is not supported for this path. Adding this context will help users understand the constraint better.

logger.warning( f"Flashinfer cutlass MoE and EP MoE are enabled. Forcing ep_size=tp_size ({self.tp_size}) because hybrid EP/TP is not supported for this path." )

…t#8630)

Fix flashinfer cutlass EP moe

5d64ae3

trevor-m requested review from merrymercy, Ying1123, zhyncs, ispobock, HaiShaw, ch-wan, BBuf and kushanam as code owners July 31, 2025 19:03

zhyncs assigned ch-wan Jul 31, 2025

gemini-code-assist bot reviewed Jul 31, 2025

View reviewed changes

zhyncs added bug Something isn't working high priority labels Jul 31, 2025

gemini-code-assist bot reviewed Jul 31, 2025

View reviewed changes

pavanimajety approved these changes Jul 31, 2025

View reviewed changes

zhyncs merged commit b7170cc into sgl-project:main Jul 31, 2025
51 of 62 checks passed

zminglei pushed a commit to zminglei/sglang that referenced this pull request Jul 31, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (sgl-projec…

807920e

…t#8630)

slin1237 pushed a commit that referenced this pull request Aug 1, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (#8630)

4e8768b

huangzhilin-hzl pushed a commit to huangzhilin-hzl/sglang that referenced this pull request Aug 1, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (sgl-projec…

31cd0e5

…t#8630)

TianQiLin666666 pushed a commit to TianQiLin666666/sglang that referenced this pull request Aug 1, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (sgl-projec…

db170bc

…t#8630)

ch-wan mentioned this pull request Aug 2, 2025

[Bug] CUDA graph error for w4afp8 ep-moe #8601

Closed

5 tasks

lifuhuang pushed a commit that referenced this pull request Aug 3, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (#8630)

702001c

ShangmingCai pushed a commit that referenced this pull request Aug 5, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (#8630)

9169258

ShangmingCai pushed a commit that referenced this pull request Aug 5, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (#8630)

f3286c6

narutolhy pushed a commit to narutolhy/sglang that referenced this pull request Aug 17, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (sgl-projec…

eafa81b

…t#8630)

narutolhy pushed a commit to narutolhy/sglang that referenced this pull request Aug 18, 2025

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor (sgl-projec…

3a6a7e3

…t#8630)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor #8630

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor #8630

Uh oh!

trevor-m commented Jul 31, 2025

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot left a comment

Uh oh!

gemini-code-assist bot Jul 31, 2025

Uh oh!

Uh oh!

Uh oh!

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor #8630

[bugfix] Fix flashinfer cutlass EP moe after MoE refactor #8630

Uh oh!

Conversation

trevor-m commented Jul 31, 2025

Motivation

Modifications

Accuracy Test

Benchmark & Profiling

Checklist

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jul 31, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!