---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#110-dspark-drafter|dsv41 §1.10]]'
  - '[[thoughts/inference/deepseek-v4.1-serving#4-pr-reading-order|dsv41 §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/csrc/libtorch_stable/torch_bindings.cpp
  - file:///Users/aarnphm/workspace/vllm/csrc/libtorch_stable/ops.h
merged: '2026-09-16'
number: 56441
source:
  - https://github.com/vllm-project/vllm/blob/6ecd97f1b7fb13991065370f6d6be3dffe9fe1db/csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu
  - https://github.com/vllm-project/vllm/blob/6ecd97f1b7fb13991065370f6d6be3dffe9fe1db/vllm/models/deepseek_v41/nvidia/dspark.py
  - https://github.com/vllm-project/vllm/blob/6ecd97f1b7fb13991065370f6d6be3dffe9fe1db/csrc/libtorch_stable/torch_bindings.cpp
  - https://github.com/vllm-project/vllm/blob/6ecd97f1b7fb13991065370f6d6be3dffe9fe1db/csrc/libtorch_stable/ops.h
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#56441 [Perf][DSpark] Add KV-only context insertion across V4.1 cache formats'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/56441
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/56441
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/56441.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Replace the DeepSeek-V4.1 DSpark context insertion call’s unused query allocation and query work with a KV-only wrapper. Dispatch the existing CUDA KV math with zero query heads for packed uint8 `fp8_ds_mla`, BF16 and ordinary FP8 e4m3fn… · full dump: [[thoughts/inference/vllm-prs/56441|#56441]]

