sglang-diffusion-add-model

To add a diffusion model to SGLang, follow a structured workflow covering reference study, component reuse, configuration creation, and registry wiring. This integration involves connecting DiT components, VAEs, text encoders, and schedulers into an end-to-end pipeline.

Hybrid and Modular pipeline styles in SGLang accommodate model-specific preprocessing or the reuse of standard stages. Choosing between them depends on whether your diffusion model requires custom integration logic or can assemble existing standard pipeline components directly.

Yes, you can reuse Diffusers pipeline components when integrating a model into SGLang. The workflow includes a reference study step to analyze existing Diffusers pipelines, allowing you to identify and repurpose standard stages like text encoders and schedulers.

The best way to wire DiT components and VAEs into a production-grade diffusion pipeline is using a step-by-step template covering stage implementation and registration. This ensures correct registry entries and valid pipeline integration within SGLang.

Yes, you need to create configuration files to integrate diffusion models into SGLang. The integration workflow explicitly includes a config creation step to define how the model's components and pipeline stages are structured and executed.

A new diffusion pipeline fails to register in SGLang if the integration workflow is incomplete. You must follow the end-to-end steps including component implementation, configuration creation, and proper registry wiring to ensure the pipeline is recognized.