z-lab/dflash: Block Diffusion for Flash Speculative Decoding Gains +671 Stars Today at 3.6K Total

GitHub·medium signal

DFlash introduces a block diffusion approach to speculative decoding that generates multiple candidate tokens in parallel for faster LLM inference. Based on the paper 'DFlash: Block Diffusion for Flash Speculative Decoding' (arXiv:2602.06036), the repo provides pre-trained draft models for Qwen, Gemma, and other popular LLMs. At +671 stars today, it's the fastest-rising inference optimization tool on GitHub this week.