---
cited_by:
  - '[[thoughts/inference/kv-layout#0-the-vocabulary-rfc-42082-lucas-wilkinson-2026-05-08|reading-map §0]]'
  - '[[thoughts/inference/kv-transfer#5-reading-order|reading-map §5]]'
  - '[[thoughts/inference/architectures#3-what-changed-in-vllms-kv-cache-manager-to-hold-these|README §3]]'
  - '[[thoughts/inference/qwen3.8-serving#3-kv-cache--state-what-lives-where-per-request|qwen4next §3]]'
  - '[[thoughts/inference/qwen3.8-runtime#7-convergence-with-deepseek-v41-and-glm-53|qwen4next §7]]'
date: '2026-11-11'
kind: rfc
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/flash_attn.py
number: 42082
source:
  - https://github.com/vllm-project/vllm/blob/4140faa4a51d42cb9618949bee28fd47682f611c/vllm/v1/attention/backends/flash_attn.py
state: open
tags:
  - vllm
  - vllm-pr
  - hetero-tp-kv-layout
title: '#42082 [RFC]: Standardize KV-cache Layouts'
topic: hetero-tp-kv-layout
url: https://github.com/vllm-project/vllm/issues/42082
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/42082
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/42082.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Right now attention backends can have semantically different and physically different KV-cache layouts. This leads to messy KV-connector code (i.e. lots of `is_mamba` and `is_mla` flags), and creates a tight coupling between KV-connector… · full dump: [[thoughts/inference/vllm-prs/42082|#42082]]

