sglang-diffusion-add-model

To add a new diffusion model to SGLang, you must provide a reference implementation, implement model components, create configs, wire pipelines, and register the model. The process requires aligning with SGLang's pipeline architecture and correct weight naming.

Hybrid integration consolidates model-specific pre-processing into a BeforeDenoisingStage while reusing standard DenoisingStage and DecodingStage. Modular integration enables composition using existing standard stages like add_standard_t2i_stages when the new model aligns with existing pipelines.

Yes, SGLang supports custom pre-processing logic through the Hybrid integration pattern. This pattern allows you to consolidate model-specific pre-processing into a BeforeDenoisingStage before utilizing the standard DenoisingStage for the remainder of the pipeline.

An integrated diffusion model outputs noise if it lacks conformance to required pipeline stages. You must ensure correct implementation of BeforeDenoisingStage, DenoisingStage, and decoding stages, alongside proper weight naming, to produce valid, non-noise outputs.

Wiring a Diffusers pipeline into SGLang requires conformance to specific stages. You must implement and wire the pipeline through the BeforeDenoisingStage, DenoisingStage, and decoding stages to ensure the pipeline architecture processes inputs correctly and generates valid outputs.