---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#19-flashmla-mega-attention-sm100|dsv41 §1.9]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-10-06'
number: 60064
source:
  - https://github.com/vllm-project/vllm/blob/9a1f6fd4cbe4ed8fd26215bc32a3f285b6a92d21/vllm/models/deepseek_v41/nvidia/flash_mla_mega_attn.py
  - https://github.com/vllm-project/vllm/blob/9a1f6fd4cbe4ed8fd26215bc32a3f285b6a92d21/vllm/models/deepseek_v41/common/ops/fused_layout.py
  - https://github.com/vllm-project/vllm/blob/9a1f6fd4cbe4ed8fd26215bc32a3f285b6a92d21/vllm/models/deepseek_v41/nvidia/dspark.py
  - https://github.com/vllm-project/vllm/blob/9a1f6fd4cbe4ed8fd26215bc32a3f285b6a92d21/vllm/models/deepseek_v41/nvidia/model.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#60064 [Model][DeepSeek-V4.1] Permute mega-attention wq_b / wo_a while loading'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/60064
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/60064
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/60064.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
DeepSeek-V4.1’s FlashMLA mega attention (the SM100 default) permutes `wq_b` rows and `wo_a` columns into the kernel’s layout as a post-load step. That step is only correct when it runs before the linear quant hooks. This PR does the… · full dump: [[thoughts/inference/vllm-prs/60064|#60064]]

