---
cited_by:
  - '[[thoughts/inference/replayssm-serving#5-pr-reading-order|replayssm §5]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/attention/backends/flashinfer.py
merged: '2026-08-13'
number: 52148
source:
  - https://github.com/vllm-project/vllm/blob/b652dedd0c505b6d300788f9b4f6c494d2d12d5c/vllm/v1/attention/backends/flashinfer.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#52148 [Attention] Fix FlashInfer SM12x prefill with sinks'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/52148
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/52148
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/52148.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Use FlashInfer’s sink-aware paged prefill wrapper on SM12x when attention sinks are enabled. The generic FA2 prefill path accepts a `sinks` argument but does not apply it, so [#49718](https://github.com/aarnphm/aarnphm.github.io/issues/49718) can use XQA for decode while producing incorrect prefill… · full dump: [[thoughts/inference/vllm-prs/52148|#52148]]

