xDiT documentation
Run image, video, and audiovisual diffusion models on one device or across multiple devices and nodes. xDiT supports CUDA, ROCm, and Ascend environments; Model and feature availability depends on the device and installed software.
Supported models · Quick start · CLI reference
Run a model
- Choose a model. Check its accepted names, parallelism, memory, precision, and caching support in the feature tables.
- Set up your environment. Install the packages or choose a container for your device and model.
- Generate an output. Start on one device, then use launching to add devices or nodes. The examples cover images, video, audio, and hardware recipes.
The CLI reference groups options by purpose and explains their values and defaults. Individual guides describe compatibility and practical use.
Scale and optimize
Start with the model's defaults and measure a baseline. Ulysses is a useful first layout for distributing attention. Choose additional optimizations for your workload, available memory, and device connections.
| Goal | Guides |
|---|---|
| Distribute inference | Ulysses and Ring, hybrid layouts, CFG, PipeFusion |
| Fit the workload in memory | Loading, sharding, and offload, parallel VAE |
| Tune kernels and precision | Attention backends, GEMM quantization, FP8 communication and Attention A2A |
| Reuse denoising work | Step caching |
| Adapt and measure | LoRA, benchmarking and outputs |
Check model support before combining features. Reduced precision and caching can change output quality; compare results and timing on your workload.
Integrate and contribute
Use xDiT ComfyUI or explore cloud, device, and container guides.
For development, start with framework design and adding a model. Contributors can build and update the docs. Look for 中文 on guides with a Chinese edition.