---
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/model_executor/layers/quantization/utils/nvfp4_emulation_utils.py
  - file:///Users/aarnphm/workspace/vllm/vllm/model_executor/layers/quantization/modelopt.py
merged: '2026-09-09'
number: 55713
source:
  - https://github.com/vllm-project/vllm/blob/83fe99399ec0603b32393a14324b65c67ad04af2/vllm/model_executor/layers/quantization/utils/nvfp4_emulation_utils.py
  - https://github.com/vllm-project/vllm/blob/83fe99399ec0603b32393a14324b65c67ad04af2/vllm/model_executor/models/qwen3_dspark.py
  - https://github.com/vllm-project/vllm/blob/83fe99399ec0603b32393a14324b65c67ad04af2/vllm/model_executor/layers/quantization/modelopt.py
state: merged
tags:
  - vllm
  - vllm-pr
  - new-architectures
title: '#55713 [Spec Decode] Add NVFP4 DSpark gathered top-k projection'
topic: new-architectures
url: https://github.com/vllm-project/vllm/pull/55713
created: '2026-11-11'
modified: '2026-11-11'
published: '2026-11-11'
pageLayout: default
slug: hinterland/papers/maps/prs/55713
permalink: https://aarnphm.xyz/hinterland/papers/maps/prs/55713.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
DSpark’s Markov head computes the draft correction by selecting rows from its W2 projection. With a quantized W2 Markov head, those rows are stored as packed NVFP4 values plus per-group/global scales, so they cannot be indexed and used as… · full dump: [[thoughts/inference/vllm-prs/55713|#55713]]

