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

Researchers at UC San Diego have introduced DFlash, a speculative decoding technique that uses a block diffusion model to predict multiple tokens simultaneously rather than one by one. This approach enables parallel generation during the drafting phase, achieving up to a 6x speed increase for models like Qwen3-8B without sacrificing output accuracy. By integrating directly with NVIDIA hardware, this method improves throughput efficiency for large language models, potentially reducing the latency bottlenecks typically associated with autoregressive generation.