---
cited_by:
  - '[[thoughts/inference/qwen3.8-runtime#14-mtp--nvidiamtppy|qwen4next §1.4]]'
  - '[[thoughts/inference/qwen3.8-serving#4-pr-reading-order-vllm-main-git-show-sha|qwen4next §4]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-09-04'
number: 54687
source:
  - https://github.com/vllm-project/vllm/blob/fd4a1512628ad17944095263c1fe89598710a3ce/vllm/models/qwen4_exp/nvidia/ops/hc.py
  - https://github.com/vllm-project/vllm/blob/fd4a1512628ad17944095263c1fe89598710a3ce/vllm/models/qwen4_exp/nvidia/hyperconnection.py
  - https://github.com/vllm-project/vllm/blob/fd4a1512628ad17944095263c1fe89598710a3ce/vllm/models/qwen4_exp/nvidia/mtp.py
  - https://github.com/vllm-project/vllm/blob/fd4a1512628ad17944095263c1fe89598710a3ce/vllm/models/qwen4_exp/nvidia/model.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#54687 [Kernel] Reuse Qwen4Exp HC combine-norm for MTP input'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/54687
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/54687
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/54687.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Reuse the existing Qwen4Exp HC combine kernels for MTP input instead of maintaining an MTP-only add-and-norm kernel. · full dump: [[thoughts/inference/vllm-prs/54687|#54687]]

