vllm-project/vllm

[Feature]: Remove attention layer name from `unified_kv_cache_update`

Open

#33267 opened on Jan 28, 2026

View on GitHub
 (7 comments) (1 reaction) (2 assignees)Python (80,034 stars) (16,816 forks)batch import
feature requestgood first issuehelp wantedstartup-uxtorch.compile

Description

🚀 The feature, motivation and pitch

unified_kv_cache_update appears in piecewise cudagraph regions, and each layer has a different name, so each of these have to be compiled separately. This increases cold start compilation time with Dynamo partition because the graphs can no longer be reused.

We should use the same approach as #32805 and #33184 to iterate through the layers in the forward context.

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
performancebackend
Issue type
feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-3 hours
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.
clear
Prerequisites
Python proficiencyUnderstanding of CUDA graphsFamiliarity with vLLM forward context
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
40
Research direction
Study the approach used in PRs #32805 and #33184, which likely iterate through layers in the forward context to avoid per layer compilation. Identify the `unified kv cache update` function in the codebase (search for its definition and usage). Look at how the forward context is structured and how layer names are currently used. Then adapt the pattern from the referenced PRs to remove the layer name dependency, possibly by collecting all layers into a single call or using a generic name.