z-lab

DFlash is a lightweight block diffusion model designed for speculative decoding that accelerates LLM inference through efficient parallel token drafting. It provides pre-trained draft models for popular LLMs like Qwen and LLaMA, enabling significant speedups in text generation.