xDiT

Serving diffusion transformers across many GPUs is the niche xDiT fills. The engine parallelizes DiT inference through several composable techniques: Unified Sequence Parallelism, PipeFusion's patch-level pipeline parallelism that exploits redundancy between adjacent diffusion steps, CFG parallel for classifier-free guidance, and conventional data and tensor parallelism, with cache-based accelerations such as TeaCache and First-Block-Cache stackable on top. More than 30 models are supported, including FLUX, HunyuanVideo, CogVideoX, SANA, PixArt, and Stable Diffusion 3, running on NVIDIA GPUs as well as AMD MI300X class accelerators, with attention backends covering Flash Attention, cuDNN, and AITER. Installation is pip install xfuser, from source, or via Docker images, and the repository documents both the framework design and the process for adding new models. The methods are described in arXiv paper 2411.01738. Apache-2.0 licensed with roughly 2,700 stars and 767 commits, it is aimed at deployments where single-GPU latency for large image and video models is unacceptable.