vllm-project/vllm

[Feature]: Use kernel abstraction for fp4 scaled mm

Open

#31823 opened on Jan 6, 2026

View on GitHub
 (12 comments) (0 reactions) (1 assignee)Python (80,034 stars) (16,816 forks)batch import
feature requesthelp wantedunstale

Description

🚀 The feature, motivation and pitch

See MPLinearKernel and ScaledMMLinearKernel.

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
python
Domain
backendmachine learningperformance
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.
over 1 week
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
blocked
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
needs investigation
Prerequisites
Familiarity with vllm's linear kernel classesUnderstanding of fp4 quantizationCUDA kernel programming
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
Examine the existing MPLinearKernel and ScaledMMLinearKernel implementations in the vllm codebase to understand the current kernel abstraction for scaled matrix multiplication. Then design a new kernel abstraction specifically for fp4 (4 bit floating point) scaled MM, considering the discussion in the issue comments (12 comments) for any design constraints or preferences. No pull requests are linked, so this is a clean design task. Key files to review: likely in vllm/model executor/layers/ or similar paths.