---
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: 54915
source:
  - https://github.com/vllm-project/vllm/blob/a5c9179e731e8d2ec7e03485a3ab0754a5a27182/vllm/models/qwen4_exp/nvidia/ops/qsa_indexer.py
  - https://github.com/vllm-project/vllm/blob/a5c9179e731e8d2ec7e03485a3ab0754a5a27182/vllm/models/qwen4_exp/common/qsa_cache.py
  - https://github.com/vllm-project/vllm/blob/a5c9179e731e8d2ec7e03485a3ab0754a5a27182/vllm/models/qwen4_exp/nvidia/indexer_qsa.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#54915 [Qwen3.8-Flash-Next] Compact indexer logits workspace to improve prefill efficiency'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/54915
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/54915
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/54915.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
During prefill, we apply chunking on indexer to cap the memory usage consumed by logits workspace. This logits workspace have shape `[num_tokens, ctx_len]` representing QK product. Currently we size `ctx_len` based on worst possible case · full dump: [[thoughts/inference/vllm-prs/54915|#54915]]

