---
cited_by:
  - '[[thoughts/inference/qwen3.8-serving#4-pr-reading-order-vllm-main-git-show-sha|qwen4next §4]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-09-04'
number: 54873
source:
  - https://github.com/vllm-project/vllm/blob/31a8a266622781917cef482d01db415cfa3cbd92/vllm/models/qwen4_exp/nvidia/ops/qsa.py
  - https://github.com/vllm-project/vllm/blob/31a8a266622781917cef482d01db415cfa3cbd92/vllm/model_executor/warmup/qwen4_exp_qsa_warmup.py
  - https://github.com/vllm-project/vllm/blob/31a8a266622781917cef482d01db415cfa3cbd92/vllm/models/qwen4_exp/nvidia/indexer_qsa.py
  - https://github.com/vllm-project/vllm/blob/31a8a266622781917cef482d01db415cfa3cbd92/vllm/models/qwen4_exp/nvidia/qsa.py
  - https://github.com/vllm-project/vllm/blob/31a8a266622781917cef482d01db415cfa3cbd92/vllm/models/qwen4_exp/nvidia/ops/qsa_indexer.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#54873 [Qwen3.8-Flash-Next] Improve QSA sparse GQA for prefill and short-ctx decode'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/54873
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/54873
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/54873.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
QSA sparse attention read topk indices from `[num_tokens, 2048+3]` buffer, where each token can attend up to 2048+3 past KV tokens. This number can be smaller when a new token’s position is less than 2048, hence the indices buffer might be… · full dump: [[thoughts/inference/vllm-prs/54873|#54873]]

