vllm-project/vllm

[Feature]: Fused Kernel for GPT-OSS Router

Open

#28986 opened on Nov 19, 2025

View on GitHub
 (10 comments) (0 reactions) (0 assignees)Python (80,034 stars) (16,816 forks)batch import
feature requestgood first issuehelp wantedstale

Description

🚀 The feature, motivation and pitch

  • Right now, we spend ~3.5% of the layer in the expert selection
  • The operation is unfused

Write a fused kernel like we have for deepseek grouped_topk

Alternatives

  • torch compile
  • triton
  • cuda

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
pythonccpp
Domain
backendperformance
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
5
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
mostly clear
Prerequisites
GPU programmingCUDA/Triton kernel developmentvLLM architectureAttention mechanism
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
8
Research direction
Examine the existing deepseek grouped topk fused kernel in the vLLM repository (likely in csrc/ or kernels/). Look at the current GPT OSS router implementation (e.g., in vllm/model executor/layers/fused moe/ or similar). Identify the unfused operations and consider how to fuse them. Review issue comments for any initial design suggestions or performance profiling data. Refer to the alternatives mentioned (Triton vs CUDA) and evaluate trade offs. Consider writing a prototype in Triton for rapid iteration.