---
cited_by:
  - '[[thoughts/inference/replayssm-serving#5-pr-reading-order|replayssm §5]]'
date: '2026-11-11'
kind: pr
local:
  - file:///Users/aarnphm/workspace/vllm/vllm/config/speculative.py
merged: '2026-08-04'
number: 49969
source:
  - https://github.com/vllm-project/vllm/blob/5789897aa40fbab6bdfcffaa9e83da64939286fb/vllm/v1/worker/gpu/spec_decode/dspark/speculator.py
  - https://github.com/vllm-project/vllm/blob/5789897aa40fbab6bdfcffaa9e83da64939286fb/vllm/config/speculative.py
  - https://github.com/vllm-project/vllm/blob/5789897aa40fbab6bdfcffaa9e83da64939286fb/vllm/model_executor/models/qwen3_dspark.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#49969 [Spec Decode] Add top-k DSpark Markov projection'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/49969
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/49969
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/49969.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
DSpark computes the base logits for all draft positions in parallel, then applies the Markov bias sequentially over the full draft vocabulary. The repeated full-vocabulary Markov projection is on the serial drafting path. · full dump: [[thoughts/inference/vllm-prs/49969|#49969]]

