DFlash: Block Diffusion for Flash Speculative Decoding

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 constrains 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. 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$\times$ lossless acceleration across a range of models and tasks, delivering up to 2.5$\times$ higher speedup than the state-of-the-art speculative decoding method EAGLE-3.

Large language models, such as chatbots and coding assistants, are powerful but often slow because they generate text one word at a time. This sequential process makes inference expensive and prevents modern GPUs from being fully used, especially when models need to produce long answers or reasoning traces. This paper introduces DFlash, a method that makes large language model generation faster without changing the final output quality. DFlash uses a lightweight diffusion model to propose a block of future words in parallel. The large model then checks these proposed words in parallel and keeps only the correct ones. To make these guesses more accurate, DFlash also uses internal information from the large model to guide the smaller draft model. Our experiments show that DFlash can significantly speed up generation across different models and tasks, including math, coding, and chat. By reducing inference time and serving cost, DFlash can make large language models more efficient, more accessible, and easier to deploy in real-world applications.