---
cited_by:
  - '[[thoughts/papers/qiu2026designqwen38nextarchitectureevaluation#11-qsa--qwen-sparse-attention|qwen4next §1.1]]'
  - '[[thoughts/inference/qwen3.8-serving#4-pr-reading-order-vllm-main-git-show-sha|qwen4next §4]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-10-05'
number: 59533
source:
  - https://github.com/vllm-project/vllm/blob/042ab0305cc4215e2c6fb215f1d6f2e65dd7d4ac/vllm/models/qwen4_exp/nvidia/low_latency_gemm.py
  - https://github.com/vllm-project/vllm/blob/042ab0305cc4215e2c6fb215f1d6f2e65dd7d4ac/vllm/models/qwen4_exp/nvidia/qsa.py
  - https://github.com/vllm-project/vllm/blob/042ab0305cc4215e2c6fb215f1d6f2e65dd7d4ac/vllm/models/qwen4_exp/nvidia/ops/qsa_prepare.py
  - https://github.com/vllm-project/vllm/blob/042ab0305cc4215e2c6fb215f1d6f2e65dd7d4ac/vllm/models/qwen4_exp/nvidia/model.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#59533 [Perf][Qwen4Exp] Merge QSA QKVG and indexer QK projections'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/59533
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/59533
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/59533.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Projection fusion. Merge Qwen4Exp QSA’s QKVG and replicated indexer Q/K projections into one GEMM. LoRA keeps separate projections with standard linear dispatch. · full dump: [[thoughts/inference/vllm-prs/59533|#59533]]

