---
cited_by:
  - '[[thoughts/inference/speculative-drafting#3-vllm-status-at-a-glance|reading-map §3]]'
date: '2026-11-11'
kind: pr
local: []
merged: '2026-09-24'
number: 57632
source:
  - https://github.com/vllm-project/vllm/blob/04730e82700d0c9b957cda188dd1bbc8073e1fcd/vllm/v1/worker/gpu/spec_decode/dflash/speculator.py
  - https://github.com/vllm-project/vllm/blob/04730e82700d0c9b957cda188dd1bbc8073e1fcd/vllm/v1/worker/gpu/spec_decode/dflash/cudagraph.py
state: merged
tags:
  - vllm
  - vllm-pr
  - specdec
title: '#57632 [DFlash] Capture the context K/V precompute in the draft CUDA graph'
topic: specdec
url: https://github.com/vllm-project/vllm/pull/57632
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/57632
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/57632.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
Even when the query forward replays a FULL CUDA graph, the DFlash/DSpark draft step still runs the context K/V precompute (`precompute_and_store_context_kv`) eagerly outside the graph, because the number of context tokens varies from step… · full dump: [[thoughts/inference/vllm-prs/57632|#57632]]

