add-diffusion-model

To add a diffusion model, you port the transformer, register it in the diffusion registry, and configure multi-GPU parallelism patterns like tensor and sequence parallelism. This workflow ensures correct model_index.json setup and Python package exports.

Yes, vLLM-Omni supports integrating both Diffusers-based pipelines and custom repositories. The integration process involves porting transformers, registering the model, and testing to ensure compatibility across different pipeline architectures.

Cache-DiT acceleration optimizes diffusion model inference within vLLM-Omni by caching intermediate computations. You integrate it during the transformer porting phase to speed up generation without altering the core model architecture.

vLLM-Omni supports tensor parallelism (TP), sequence parallelism (SP/USP), classifier-free guidance (CFG), and hybrid sharded data parallelism (HSDP). These patterns enable scalable, production-ready multi-GPU deployment for diffusion models.

You must register the model in the diffusion registry, ensure proper Python package exports, and run an offline test script. Scaffolding offline download scripts and updating production documentation are also required for complete integration.

Use sequence parallelism (SP/USP) when scaling diffusion models across multiple GPUs to handle long sequences efficiently, whereas tensor parallelism (TP) partitions model weights. vLLM-Omni supports both, often combining them with CFG and HSDP.