vllm-project/vllm

[Feature]: Upstream DGX spark improvements from Avarok-Cybersecurity/dgx-vllm

Open

#37141 opened on Mar 16, 2026

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

Description

🚀 The feature, motivation and pitch

As shown in the readme in Avarok-Cybersecurity/dgx-vllm, there are gaps in vLLM fp4 performance for DGX. The fixes seem not too complicated, and we should try to upstream the changes from the repo into vLLM.

Avarok-Cybersecurity/dgx-vllm#7

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
performance
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.
needs investigation
Prerequisites
familiarity with vLLM codebaseunderstanding of fp4 precisionexperience with GPU kernel optimization
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
20
Research direction
Start by reviewing the referenced repository Avarok Cybersecurity/dgx vllm and its issue #7 to understand the specific improvements made for fp4 performance on DGX. Since no code diff is provided in this issue, the next step is to clone the fork, identify the relevant commits or changes, and assess how they could be integrated into the main vLLM codebase. Note that this is a complex task involving low level performance optimizations, and it lacks a clear list of deliverables. Maintainers should clarify which parts of the fork are intended to be upstreamed and the expected scope of work.