vllm-project/vllm

[Feature]: Extract KV-Cache update from all attention backends

Open

#32,335 建立於 2026年1月14日

在 GitHub 查看
 (51 留言) (0 反應) (9 負責人)Python (80,034 star) (16,816 fork)batch import
feature requestgood first issuehelp wanted

描述

🚀 The feature, motivation and pitch

Similar to how https://github.com/vllm-project/vllm/pull/25954 extracts it from FlashAttn. Ideally, we want to cover all backends with kv-cache update from v1/attention/backends.

Backends:

  • FlashAttention
  • FlashInfer
  • AiterFlashAttention (in progress)
  • RocmAiterUnifiedAttention
  • RocmAttention
  • TritonAttention
  • FlashAttentionDiffKV
  • FlexAttention
  • TreeAttention

MLA Backends:

  • FlashAttnMLA
  • FlashInferMLA
  • FlashMLASparse
  • FlashMLA
  • AiterMLA
  • ROCMAiterMLASparse
  • CutlassMLA
  • TritonMLA

After all backends are supported, we can remove slot_mapping from attention metadata.

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.

貢獻者指南

技術棧
pythonpytorchcpp
領域
backendperformanceai
議題類型
feature
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
5
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
over 1 week
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
active
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
clear
前置要求
vllm attention backendsKV cache optimizationCUDA programmingv1/attention module
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
10
研究方向
Review PR #25954 which demonstrates the extraction pattern for FlashAttn. Examine the existing completed backends (e.g., FlashInfer, TritonAttention) in `v1/attention/backends` to understand the structure. The AiterFlashAttention backend is in progress; coordinate with the current assignees. The goal is to implement the KV cache update extraction for the remaining backends and eventually remove `slot mapping` from attention metadata.
[Feature]: Extract KV-Cache update from all attention backends · vllm-project/vllm#32335 | Good First Issue