---
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/attention/backends/mla/flashmla_sparse.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py
merged: '2026-09-12'
number: 53781
source:
  - https://github.com/vllm-project/vllm/blob/d43bb2f37f87a63d3a9a299c97050443cf6a2520/csrc/libtorch_stable/hisparse_kernels.cu
  - https://github.com/vllm-project/vllm/blob/d43bb2f37f87a63d3a9a299c97050443cf6a2520/vllm/distributed/kv_transfer/kv_connector/v1/hisparse/worker.py
  - https://github.com/vllm-project/vllm/blob/d43bb2f37f87a63d3a9a299c97050443cf6a2520/vllm/v1/attention/backends/mla/index_group.py
  - https://github.com/vllm-project/vllm/blob/d43bb2f37f87a63d3a9a299c97050443cf6a2520/vllm/distributed/kv_transfer/kv_connector/v1/nixl/base_worker.py
  - https://github.com/vllm-project/vllm/blob/d43bb2f37f87a63d3a9a299c97050443cf6a2520/vllm/distributed/kv_transfer/kv_connector/v1/hisparse/connector.py
  - https://github.com/vllm-project/vllm/blob/d43bb2f37f87a63d3a9a299c97050443cf6a2520/vllm/v1/attention/backends/mla/flashmla_sparse.py
  - https://github.com/vllm-project/vllm/blob/d43bb2f37f87a63d3a9a299c97050443cf6a2520/docs/design/hisparse.md
  - https://github.com/vllm-project/vllm/blob/d43bb2f37f87a63d3a9a299c97050443cf6a2520/vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#53781 [3/N] HiSparse: host-resident sparse-MLA decode hot-buffering'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/53781
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/53781
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/53781.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
HiSparse adds a host-resident tier beneath vLLM’s normal GPU KV cache for sparse-MLA decode. Unlike the original always-host-resident design in [#46326](https://github.com/aarnphm/aarnphm.github.io/issues/46326), this implementation keeps as much KV as possible in the normal device cache and spills… · full dump: [[thoughts/inference/vllm-prs/53781|#53781]]

