add-new-diffusion-model

To add a new diffusion model to SGLang, you implement model-specific pre-processing components, wire them into standard denoising and decoding stages, and register the pipeline configuration in the global registry.

The BeforeDenoisingStage is a pre-processing step in a diffusion model pipeline used for complex, model-specific data preparation before the standard denoising stage executes.

You need a modular composition style for diffusion model integration when you want to reuse existing pipeline components and stages rather than building a single monolithic pipeline from scratch.

You register a diffusion model pipeline in the global registry by wiring your model components together, defining the pipeline configuration, and adding the required registry entries for production deployment.

The best way to test a new diffusion model integration is to follow the step-by-step testing documentation after completing pipeline configuration, component wiring, and registry entry to ensure robust production deployment.