Fast-dLLM: Training-free Acceleration of Diffusion LLM by Enabling KV Cache and Parallel Decoding Paper โข 2505.22618 โข Published May 28, 2025 โข 45