DFlash: Block Diffusion for Flash Speculative Decoding
Diffusion LLM
Hello folks,
Speculative decoding is meant to fix the sequential bottleneck in autoregressive LLMs — but the drafter is usually autoregressive too. The cure inherits the disease:
1. Drafting stays sequential, which caps the practical speedup
2. Diffusion LLMs can draft in parallel, but on their own they still underperform autoregressive models
The insight: speculative decoding is exactly the setting where a diffusion model shines. The target model verifies every draft, so the drafter's parallelism is pure upside — and the output stays lossless.
DFlash uses a lightweight block diffusion model as the drafter, with two pieces:
1. Draft tokens are produced in a single forward pass — drafting stops being sequential
2. The drafter is conditioned on context features extracted from the target model, which lifts draft quality and acceptance rates
Over 6× lossless acceleration across a range of models and tasks — and up to 2.5× higher speedup than EAGLE-3, the state-of-the-art speculative decoding method.
This Monday, the lead author Jian Chen will present DFlash.
Title: DFlash: Block Diffusion for Flash Speculative Decoding
Meeting Link: click here
Time: Sep 21 (Monday) 1pm ET / 10am PT / 7pm CET / 10:30pm IST
Paper: [2602.06036] DFlash: Block Diffusion for Flash Speculative Decoding
Prior knowledge:
Fundamentals of discrete diffusion (video by Sasha Rush)
Abstract
Autoregressive large language models (LLMs) deliver strong performance but require inherently sequential decoding, leading to high inference latency and poor GPU utilization. Speculative decoding mitigates this bottleneck by using a fast draft model whose outputs are verified in parallel by the target LLM; however, existing methods still rely on autoregressive drafting, which remains sequential and limits practical speedups. Diffusion LLMs offer a promising alternative by enabling parallel generation, but current diffusion models typically underperform compared with autoregressive models. In this paper, we introduce DFlash, a speculative decoding framework that employs a lightweight block diffusion model for parallel drafting. By generating draft tokens in a single forward pass and conditioning the draft model on context features extracted from the target model, DFlash enables efficient drafting with high-quality outputs and higher acceptance rates. Experiments show that DFlash achieves over 6x lossless acceleration across a range of models and tasks, delivering up to 2.5x higher speedup than the state-of-the-art speculative decoding method EAGLE-3.
Yours truly,
Subham, Justin, Zhihan
Diffusion LLM
Meeting Link: click here
Time: Sep 21 (Monday) 1pm ET / 10am PT / 7pm CET / 10:30pm IST
Paper: [2602.06036] DFlash: Block Diffusion for Flash Speculative Decoding