---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#18-moe--gate|dsv41 §1.8]]'
  - '[[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/.buildkite/test_areas/kernels.yaml
merged: '2026-09-17'
number: 56266
source:
  - https://github.com/vllm-project/vllm/blob/41f9104fa649aed2c26af8a459210619ad9514fe/vllm/models/deepseek_v4/nvidia/model.py
  - https://github.com/vllm-project/vllm/blob/41f9104fa649aed2c26af8a459210619ad9514fe/vllm/utils/deep_gemm.py
  - https://github.com/vllm-project/vllm/blob/41f9104fa649aed2c26af8a459210619ad9514fe/vllm/models/deepseek_v4/nvidia/mtp.py
  - https://github.com/vllm-project/vllm/blob/41f9104fa649aed2c26af8a459210619ad9514fe/vllm/models/deepseek_v41/nvidia/model.py
  - https://github.com/vllm-project/vllm/blob/41f9104fa649aed2c26af8a459210619ad9514fe/vllm/models/deepseek_v4/nvidia/dspark.py
  - https://github.com/vllm-project/vllm/blob/41f9104fa649aed2c26af8a459210619ad9514fe/vllm/models/deepseek_v41/nvidia/dspark.py
  - https://github.com/vllm-project/vllm/blob/41f9104fa649aed2c26af8a459210619ad9514fe/.buildkite/test_areas/kernels.yaml
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#56266 [DSv4.1] Integrate Mega-Gate from DeepGEMM'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/56266
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/56266
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/56266.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Integrate DeepGEMM Mega-Gate into the DeepSeek V4/V4.1 DeepGEMM MegaMoE path, fusing the gate GEMM and expert selection. Prepare routing metadata once per model forward and reuse it across layers. Preserve hash routing for V4, image… · full dump: [[thoughts/inference/vllm-prs/56266|#56266]]

