Papers & Research

DFlash: block diffusion draft model delivers up to 15x throughput gain

Original title:DFlash Speculative Decoding Drafts Whole Token Blocks in Parallel for Up to 15x Higher Throughput on NVIDIA Blackwell

AI summary

DFlash, proposed by the UC San Diego team, is a lightweight block diffusion draft model for speculative decoding. It generates an entire block of tokens in one forward pass, which the target model then verifies in parallel, guaranteeing lossless output. Compared with EAGLE-3, DFlash achieves up to 2.5x speedup and averages over 6x lossless speedup across models such as Qwen3-8B (6.08x on MATH-500). On NVIDIA Blackwell (TensorRT-LLM), the gpt-oss-120b model sees up to 15x throughput improvement, about 1.5x that of EAGLE-3. The core innovation injects the target model's multi-layer hidden features into the Key-Value projection of each draft model layer, so acceptance length grows with draft depth.

Key Takeaways

  • DFlash,proposed by the UC San Diego team,is a lightweight block diffusion draft model for…
  • It generates an entire block of tokens in one forward pass,which the target model then verifies in…
  • Compared with EAGLE-3,DFlash achieves up to 2.
  • 5x speedup and averages over 6x lossless speedup across models such as Qwen3-8B (6.
  • 08x on MATH-500).

Quick Facts

Published
Updated
Source
MarkTechPost(RSS)
Category
Papers & Research
AI score
74
Read the original Read the full article at the original source
Why it matters

DFlash把扩散模型引入推测解码草案阶段,一次并行生成整块token,单流加速最高6倍,NVIDIA实测吞吐量提升15倍,推理成本下降幅度很大,部署大模型的团队可以立刻关注。

Related entities

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: DFlash: block diffusion draft model delivers up to 15x throughput gain. June 24, 2026. https://new.nat6.net/en/news/cmqrrgnay0n08slp5phvdk394/

View all →