---
cited_by:
  - '[[thoughts/inference/deepseek-v4.1-serving#5-sharp-edges-bugs-and-fixes-that-expose-invariants|dsv41 §5]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-08-10'
number: 51430
source:
  - https://github.com/vllm-project/vllm/blob/79c865b838e34f7a98a936771284773819d79c8f/vllm/models/deepseek_v4/attention.py
  - https://github.com/vllm-project/vllm/blob/79c865b838e34f7a98a936771284773819d79c8f/vllm/models/deepseek_v4/amd/model.py
  - https://github.com/vllm-project/vllm/blob/79c865b838e34f7a98a936771284773819d79c8f/vllm/models/deepseek_v4/nvidia/model.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#51430 [Perf] Narrow DeepSeek V4 eager CUDA graph region'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/51430
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/51430
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/51430.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Narrow the DeepSeek V4 eager CUDA graph region so attention input preparation remains captured, and organize the flow so the execution stages are visible in one linear `forward` method. · full dump: [[thoughts/inference/vllm-prs/51430|#51430]]

