---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#110-dspark-drafter|dsv41 §1.10]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/mla/sparse_swa.py
merged: '2026-09-17'
number: 57432
source:
  - https://github.com/vllm-project/vllm/blob/80447d27655918da6bfbccd0d3a40e975bda220a/vllm/models/deepseek_v41/nvidia/flashinfer_sparse.py
  - https://github.com/vllm-project/vllm/blob/80447d27655918da6bfbccd0d3a40e975bda220a/vllm/v1/attention/backends/mla/sparse_swa.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#57432 [Bugfix][DSv4.1] Fix FlashInfer DSpark non-causal attention'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/57432
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/57432
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/57432.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
DSpark’s non-causal draft window has 133 valid keys at draft length 5, padded to 256 indices. The DSV4.1 FlashInfer integration passed the padded width as the active sparse length, so padding entered softmax normalization. A constant-value… · full dump: [[thoughts/inference/vllm-prs/57432|#57432]]

