---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#4-pr-reading-order|dsv41 §4]]'
  - '[[thoughts/inference/agent-review#41-fast-track-rfc-59665|reading-map §4.1]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-10-10'
number: 60397
source:
  - https://github.com/vllm-project/vllm/blob/193922d6ff5f7e213eacadd3cb7dd96612b30ed3/vllm/models/deepseek_v41/amd/mono/stages/moe.py
  - https://github.com/vllm-project/vllm/blob/193922d6ff5f7e213eacadd3cb7dd96612b30ed3/vllm/models/deepseek_v41/amd/mono/layer.py
  - https://github.com/vllm-project/vllm/blob/193922d6ff5f7e213eacadd3cb7dd96612b30ed3/vllm/models/deepseek_v41/amd/mono/attention/back.py
  - https://github.com/vllm-project/vllm/blob/193922d6ff5f7e213eacadd3cb7dd96612b30ed3/vllm/models/deepseek_v41/amd/mono/attention/front.py
  - https://github.com/vllm-project/vllm/blob/193922d6ff5f7e213eacadd3cb7dd96612b30ed3/vllm/models/deepseek_v41/amd/mono/attention/device.py
  - https://github.com/vllm-project/vllm/blob/193922d6ff5f7e213eacadd3cb7dd96612b30ed3/vllm/models/deepseek_v41/amd/mono/runner.py
  - https://github.com/vllm-project/vllm/blob/193922d6ff5f7e213eacadd3cb7dd96612b30ed3/vllm/models/deepseek_v41/amd/mono/stages/seam.py
  - https://github.com/vllm-project/vllm/blob/193922d6ff5f7e213eacadd3cb7dd96612b30ed3/vllm/models/deepseek_v41/amd/mono_decode.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#60397 [ROCm][Perf][DSv4.1] Mono decode layer for DeepSeek-V4.1-Flash on MI355X'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/60397
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/60397
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/60397.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
All TP all-reduces run in-kernel over symmetric peer memory. K1 + K2 replace the \~40 kernels this path launches per standard layer today; the FFN launch replaces the attention all-reduce, the FFN seam and the MoE’s kernels on the others. · full dump: [[thoughts/inference/vllm-prs/60397|#60397]]

