---
cited_by:
  - '[[thoughts/inference/dflash2|reading-map §2.2]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/.buildkite/test_areas/spec_decode.yaml
merged: '2026-09-17'
number: 55960
source:
  - https://github.com/vllm-project/vllm/blob/db7a24c230a4db6ae3a568ed01b01f3f7172069d/vllm/model_executor/models/qwen3_dflash2.py
  - https://github.com/vllm-project/vllm/blob/db7a24c230a4db6ae3a568ed01b01f3f7172069d/.buildkite/test_areas/spec_decode.yaml
state: merged
tags:
  - vllm
  - vllm-pr
  - specdec
title: '#55960 [Perf] Add fused DFlash2 grouped convolution'
topic: specdec
url: https://github.com/vllm-project/vllm/pull/55960
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/55960
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/55960.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Fuse DFlash2’s grouped dynamic convolution in one Triton custom op, keeping the existing Torch CPU fallback. The kernel stays in `qwen3_dflash2.py`. · full dump: [[thoughts/inference/vllm-prs/55960|#55960]]

