---
cited_by:
  - '[[thoughts/inference/architectures#4-pd-and-hetero-tp-per-model|README §4]]'
  - '[[thoughts/inference/architectures#6-reading-order|README §6]]'
  - '[[thoughts/inference/deepseek-v4.1-state#35-pd-and-offload|dsv41 §3.5]]'
  - '[[thoughts/inference/deepseek-v4.1-serving#4-pr-reading-order|dsv41 §4]]'
  - '[[thoughts/inference/replayssm#4-which-models-use-it|replayssm §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/core/sched/scheduler.py
  - file:///Users/aarnphm/workspace/vllm/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_connector.py
merged: '2026-10-07'
number: 59197
source:
  - https://github.com/vllm-project/vllm/blob/ef63c23d35acccbba8e014fc88da19d541da38a9/vllm/v1/core/sched/scheduler.py
  - https://github.com/vllm-project/vllm/blob/ef63c23d35acccbba8e014fc88da19d541da38a9/vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_connector.py
  - https://github.com/vllm-project/vllm/blob/ef63c23d35acccbba8e014fc88da19d541da38a9/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_scheduler.py
  - https://github.com/vllm-project/vllm/blob/ef63c23d35acccbba8e014fc88da19d541da38a9/vllm/distributed/aux_output_connector/connector.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#59197 [Bugfix][DSv4.1] Skip SWA bounded replay for KV loads that carry the window'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/59197
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/59197
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/59197.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
SWA bounded replay ([#56227](https://github.com/aarnphm/aarnphm.github.io/issues/56227)) replays the last window of every prefix hit, including KV loaded by P/D connectors, which transfer the request’s own sliding-window KV. P/D decode on DeepSeek-V4.1 therefore recomputed the last 128 prompt tokens… · full dump: [[thoughts/inference/vllm-prs/59197|#59197]]

