Here's the abstract from the paper. Make of that what you will: 

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 us- ing 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 constrains practi- cal 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. We show that speculative decoding provides a natural and effective setting for diffusion models. 

By generating draft tokens in a single forward pass, DFlash enables efficient drafting, and by conditioning the draft model on context features extracted from the target model, it achieves high-quality drafts with higher acceptance rates.

Experiments show that DFlash achieves over 6× lossless acceleration across a range of models and tasks, delivering up to 2.5× higher speedup than the state-of the-art speculative decoding method EAGLE-3.