---
cited_by:
  - '[[thoughts/inference/architectures#3-what-changed-in-vllms-kv-cache-manager-to-hold-these|README §3]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/kv_offload/cpu/shared_offload_region.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/kv_cache_interface.py
merged: '2026-09-12'
number: 56629
source:
  - https://github.com/vllm-project/vllm/blob/e19a3e172ecc1d1ece8ecce9dfd91bf811a814a0/vllm/v1/hisparse/runtime.py
  - https://github.com/vllm-project/vllm/blob/e19a3e172ecc1d1ece8ecce9dfd91bf811a814a0/vllm/v1/kv_offload/cpu/shared_offload_region.py
  - https://github.com/vllm-project/vllm/blob/e19a3e172ecc1d1ece8ecce9dfd91bf811a814a0/vllm/v1/hisparse/binding.py
  - https://github.com/vllm-project/vllm/blob/e19a3e172ecc1d1ece8ecce9dfd91bf811a814a0/vllm/distributed/kv_transfer/kv_connector/v1/hisparse/worker.py
  - https://github.com/vllm-project/vllm/blob/e19a3e172ecc1d1ece8ecce9dfd91bf811a814a0/vllm/v1/hisparse/layout.py
  - https://github.com/vllm-project/vllm/blob/e19a3e172ecc1d1ece8ecce9dfd91bf811a814a0/vllm/distributed/kv_transfer/kv_connector/v1/offloading/config.py
  - https://github.com/vllm-project/vllm/blob/e19a3e172ecc1d1ece8ecce9dfd91bf811a814a0/docs/design/hisparse.md
  - https://github.com/vllm-project/vllm/blob/e19a3e172ecc1d1ece8ecce9dfd91bf811a814a0/vllm/v1/kv_cache_interface.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#56629 [5/N] Share HiSparse host cache across TP ranks (reopens #52760)'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/56629
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/56629
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/56629.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Share the replicated HiSparse MLA host cache across local TP ranks using one mmap-backed pool. TP rank 0 writes the shared pool; peers wait on its IPC events before reading it. Other executor and parallel layouts retain private pools. · full dump: [[thoughts/inference/vllm-prs/56629|#56629]]

