---
cited_by:
  - '[[thoughts/inference/dflash2|reading-map §2.2]]'
  - '[[thoughts/inference/speculative-drafting#3-vllm-status-at-a-glance|reading-map §3]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/worker/gpu/sample/gumbel.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/logits_processor.py
  - file:///Users/aarnphm/workspace/vllm/vllm/config/vllm.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/models/qwen3_dflash.py
  - file:///Users/aarnphm/workspace/vllm/vllm/v1/worker/gpu/spec_decode/__init__.py
merged: '2026-08-21'
number: 52816
source:
  - https://github.com/vllm-project/vllm/blob/b389ac29465b33f9e9c534df221ea3c129e9793f/vllm/model_executor/models/qwen3_dflash2.py
  - https://github.com/vllm-project/vllm/blob/b389ac29465b33f9e9c534df221ea3c129e9793f/vllm/v1/worker/gpu/spec_decode/dflash2/speculator.py
  - https://github.com/vllm-project/vllm/blob/b389ac29465b33f9e9c534df221ea3c129e9793f/vllm/v1/worker/gpu/sample/gumbel.py
  - https://github.com/vllm-project/vllm/blob/b389ac29465b33f9e9c534df221ea3c129e9793f/vllm/model_executor/layers/logits_processor.py
  - https://github.com/vllm-project/vllm/blob/b389ac29465b33f9e9c534df221ea3c129e9793f/vllm/v1/worker/gpu/spec_decode/speculator.py
  - https://github.com/vllm-project/vllm/blob/b389ac29465b33f9e9c534df221ea3c129e9793f/vllm/config/vllm.py
  - https://github.com/vllm-project/vllm/blob/b389ac29465b33f9e9c534df221ea3c129e9793f/vllm/model_executor/models/qwen3_dflash.py
  - https://github.com/vllm-project/vllm/blob/b389ac29465b33f9e9c534df221ea3c129e9793f/vllm/v1/worker/gpu/spec_decode/__init__.py
state: merged
tags:
  - vllm
  - vllm-pr
  - specdec
title: '#52816 [Spec Decode] DFlash2: local convolution + candidate selector'
topic: specdec
url: https://github.com/vllm-project/vllm/pull/52816
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/52816
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/52816.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Each component timed alone in its own CUDA graph at the shapes the draft runs · full dump: [[thoughts/inference/vllm-prs/52816|#52816]]

