---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#4-pr-reading-order|dsv41 §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/utils/multi_stream_utils.py
merged: '2026-09-27'
number: 57407
source:
  - https://github.com/vllm-project/vllm/blob/44af287ebe38d6dc4e102948025f5e3e175aefd6/vllm/models/deepseek_v41/amd/rocm.py
  - https://github.com/vllm-project/vllm/blob/44af287ebe38d6dc4e102948025f5e3e175aefd6/vllm/models/deepseek_v41/attention.py
  - https://github.com/vllm-project/vllm/blob/44af287ebe38d6dc4e102948025f5e3e175aefd6/vllm/utils/multi_stream_utils.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#57407 [ROCm][Perf] Enable layer-aware CSA2 multi-stream overlap for DeepSeek-V4.1-Flash'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/57407
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/57407
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/57407.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
DeepSeek-V4.1-Flash has three attention layer shapes: full CSA2 layers (2/8/14/20) run SWA + compressor + indexer and build a new global KV cache with their own Top-K indices; reindex CSA2 layers (24/28/32/36) run SWA + indexer only… · full dump: [[thoughts/inference/vllm-prs/57407|#57407]]

