---
cited_by:
  - '[[thoughts/papers/xie2026mhcmanifoldconstrainedhyperconnections#16-mhc-manifold-constrained-hyper-connections-with-v41s-delayedsingle-pass-pre|dsv41 §1.6]]'
  - '[[thoughts/inference/deepseek-v4.1-serving#4-pr-reading-order|dsv41 §4]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/CMakeLists.txt
merged: '2026-09-26'
number: 58586
source:
  - https://github.com/vllm-project/vllm/blob/77871126f9b69cff9feffff390d1171a60dbd7e2/vllm/models/deepseek_v41/nvidia/ops/cute_dsl/all_reduce_mhc.py
  - https://github.com/vllm-project/vllm/blob/77871126f9b69cff9feffff390d1171a60dbd7e2/vllm/models/deepseek_v41/nvidia/ops/cute_dsl/primitives.py
  - https://github.com/vllm-project/vllm/blob/77871126f9b69cff9feffff390d1171a60dbd7e2/csrc/libtorch_stable/all_reduce_mhc.cu
  - https://github.com/vllm-project/vllm/blob/77871126f9b69cff9feffff390d1171a60dbd7e2/vllm/models/deepseek_v41/nvidia/ops/mhc.py
  - https://github.com/vllm-project/vllm/blob/77871126f9b69cff9feffff390d1171a60dbd7e2/vllm/models/deepseek_v41/nvidia/model.py
  - https://github.com/vllm-project/vllm/blob/77871126f9b69cff9feffff390d1171a60dbd7e2/vllm/models/deepseek_v41/nvidia/ops/cute_dsl/__init__.py
  - https://github.com/vllm-project/vllm/blob/77871126f9b69cff9feffff390d1171a60dbd7e2/CMakeLists.txt
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#58586 [Kernel][DSV4.1] Fuse MoE finalize into the TP all-reduce + mHC boundary'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/58586
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/58586
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/58586.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
`_SharedOnlyPublishDeviceKernel` and `_group_leader_block_sum` are byte-identical to upstream. · full dump: [[thoughts/inference/vllm-prs/58586|#58586]]

