---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#5-sharp-edges-bugs-and-fixes-that-expose-invariants|dsv41 §5]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-09-18'
number: 57454
source:
  - https://github.com/vllm-project/vllm/blob/64563d0ec4b761745d5babbc5750b263f0783926/vllm/model_executor/kernels/attention/dsa/candidate_blocks.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#57454 [Bugfix][DSv4.1] Preserve NaN-scored candidate block indices'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/57454
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/57454
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/57454.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
DeepSeek-V4.1 TP4 MegaAttention can fail startup with a DeepGEMM sparse-slot assertion during dummy prefill/autotuning. The candidate-source logits contain NaNs. Although `torch.topk` selects their block IDs, `value > -inf` discards them… · full dump: [[thoughts/inference/vllm-prs/57454|#57454]]

