vllm-project/vllm

[Feature]: Support `Phi4Flash` model in V1

Open

#23,957 opened on Aug 29, 2025

View on GitHub
 (6 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

We will drop support for V0 in the very near future, and if we want to continue supporting this model in vLLM, it needs to be ported to V1.

The main work here is in:

  1. Adapting the custom Mamba layer to match how V1 manages the mamba state. The MambaMixer, MambaMixer2, MinimaxLinearAttention and ShortConv can all be used as a reference for how to do this.
  2. Porting the differential attention backend to V1.

Alternatives

Drop model support in vLLM

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
pythonpytorch
Domain
machine learningbackend
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.
fresh
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
clear
Prerequisites
Understanding of vLLM V1 architectureFamiliarity with Mamba layersKnowledge of differential attention mechanisms
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
Examine the custom Mamba layer in the current V0 codebase and adapt it to match how V1 manages mamba state, using MambaMixer, MambaMixer2, MinimaxLinearAttention, and ShortConv as references. Then port the differential attention backend to V1. Review any related pull requests or comments in the issue for additional guidance on the V1 state management.