Fast-dLLM-mlx
An MLX port of training-free inference speedups for Dream and DiffusionGemma diffusion language models on Apple Silicon, with dual-cache decoding and benchmark scripts.
Fast-dLLM-mlx implements Dream architecture inference in MLX for Apple Silicon. It adapts the NVLabs Fast-dLLM approach to Dream-style diffusion language models, focusing on training-free inference speedups rather than changes to the underlying model.
The current implementation includes Dream architecture inference in MLX, a first MLX version of the Fast-dLLM approach, dual-cache support for more efficient decoding, and parallel token generation with probability thresholding. It also now includes classic DiffusionGemma MLX inference and a Fast-dLLM-style DiffusionGemma path for comparing block-diffusion generation on Apple Silicon, plus small benchmark scripts for comparing MLX variants across a fixed set of prompts.
Built with Python, installed with uv or pip. Apache-2.0 licensed.