---
cited_by:
  - '[[thoughts/inference/architectures#5-four-things-called-replay|README §5]]'
  - '[[thoughts/inference/architectures#6-reading-order|README §6]]'
  - '[[thoughts/inference/deepseek-v4.1-state#33-decoder-side-bounded-replay-58132-59532-59894|dsv41 §3.3]]'
  - '[[thoughts/inference/deepseek-v4.1-serving#4-pr-reading-order|dsv41 §4]]'
  - '[[thoughts/inference/review-deepseek#grill-questions-answer-key-follows-each|grill-key § grill questions (answer key…]]'
  - '[[thoughts/inference/replayssm#4-which-models-use-it|replayssm §4]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-10-09'
number: 59894
source:
  - https://github.com/vllm-project/vllm/blob/ad73a4740dd780c5620099261738a30b979b262c/vllm/models/deepseek_v41/nvidia/model_state.py
  - https://github.com/vllm-project/vllm/blob/ad73a4740dd780c5620099261738a30b979b262c/vllm/models/deepseek_v41/nvidia/model.py
  - https://github.com/vllm-project/vllm/blob/ad73a4740dd780c5620099261738a30b979b262c/vllm/models/deepseek_v41/decoder_replay_layers.py
  - https://github.com/vllm-project/vllm/blob/ad73a4740dd780c5620099261738a30b979b262c/vllm/models/deepseek_v41/attention.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#59894 [Perf][DSv4.1] Trim the last KV-source layer in decoder replay'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/59894
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/59894
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/59894.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
In decoder replay steps, the last KV-source layer (layer 20 on DSV4.1-Flash) still ran on every row, though for the trimmed rows only its KV (SWA, compressed, index K) is read. Its indexer scoring and top-k, sparse MLA, wo and MoE feed… · full dump: [[thoughts/inference/vllm-prs/59894|#59894]]

