---
cited_by:
  - '[[thoughts/papers/deepseekai2026deepseekv41flashpushinglimitskv#11-csa2-topology-per-layer-ratios-cross-layer-kv-reuse|dsv41 §1.1]]'
  - '[[thoughts/inference/deepseek-v4.1-serving#4-pr-reading-order|dsv41 §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/utils/deep_gemm.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/kv_cache_interface.py
merged: '2026-09-15'
number: 56254
source:
  - https://github.com/vllm-project/vllm/blob/12575123059c2142e39c9af4e217a4792c3c8176/vllm/model_executor/kernels/attention/dsa/sparse_mqa_logits.py
  - https://github.com/vllm-project/vllm/blob/12575123059c2142e39c9af4e217a4792c3c8176/vllm/v1/attention/backends/mla/sparse_indexer.py
  - https://github.com/vllm-project/vllm/blob/12575123059c2142e39c9af4e217a4792c3c8176/vllm/model_executor/layers/sparse_mqa_indexer.py
  - https://github.com/vllm-project/vllm/blob/12575123059c2142e39c9af4e217a4792c3c8176/vllm/utils/deep_gemm.py
  - https://github.com/vllm-project/vllm/blob/12575123059c2142e39c9af4e217a4792c3c8176/vllm/models/deepseek_v41/attention.py
  - https://github.com/vllm-project/vllm/blob/12575123059c2142e39c9af4e217a4792c3c8176/vllm/models/deepseek_v4/common/ops/fused_indexer_q.py
  - https://github.com/vllm-project/vllm/blob/12575123059c2142e39c9af4e217a4792c3c8176/vllm/models/deepseek_v4/nvidia/ops/fused_indexer_q_cutedsl.py
  - https://github.com/vllm-project/vllm/blob/12575123059c2142e39c9af4e217a4792c3c8176/vllm/v1/kv_cache_interface.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#56254 [DSA] Wire DeepGEMM sparse MQA logits into the DeepSeek V4.1 indexer'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/56254
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/56254
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/56254.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
DeepSeek V4.1 selects attention tokens in two levels: one indexer layer publishes the top candidate blocks of the context, and the indexer layers after it pick their top-k inside those blocks. Today the consumer layers still score the… · full dump: [[thoughts/inference/vllm-prs/56254|#56254]]

