Fast-DLLM: Training-Free Acceleration of Diffusion LLM (arxiv.org)

Wait, under everything I’ve read about Diffusion Language Models and demos I’ve also seen and tried, inference is faster than traditional architectures. They state the opposite what gives?