vllm-project/vllm

[Feature]: Integrate fused `kMoEFinalizeARResidualRMSNorm` from FlashInfer

Open

#40544 opened on Apr 21, 2026

View on GitHub
 (3 comments) (1 reaction) (0 assignees)Python (80,034 stars) (16,816 forks)batch import
feature requesthelp wanted

Description

🚀 The feature, motivation and pitch

Available today via FlashInfer 0.6.8, we can leverage this fused MoE Finalize + ResidualAdd + AllReduce + RMSNorm https://github.com/flashinfer-ai/flashinfer/pull/2982.

This should be added as a new torch.compile custom pass. I think moe_finalize might still be inside the wrapped fused_moe op so this might require pulling that out.

Alternatives

No response

Additional context

No response

Before submitting a new issue...

  • Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

Contributor guide

Tech stack
pythonpytorchcpp
Domain
backendperformance
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
Understanding of Mixture of Experts (MoE) modelsFamiliarity with PyTorch custom opsKnowledge of FlashInfer libraryC++/CUDA programming experience
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
Begin by studying FlashInfer PR #2982 to understand the fused kernel's interface and semantics. Then examine vLLM's MoE implementation, specifically the `fused moe` op and how `moe finalize` is currently used. The issue suggests that `moe finalize` may be embedded within the wrapped op, so you'll need to locate the relevant code in vLLM's source (likely in `csrc/` or `vllm/model executor/layers/fused moe/`). After understanding the separation, design a new torch.compile custom pass that integrates the FlashInfer kernel, ensuring it handles the finalize + residual add + all reduce + RMSNorm sequence efficiently. Consult the maintainers on the issue to clarify if any intermediate steps or modifications to existing ops are required.