RoadMap and Looking for Contributions

[feifeibear]

Model Zoo (we generally first implement USP and then PipeFusion for a new model)

• HunyuanVideo

Scheduler

• Decouple VAE and DiT backbone. They can have different parallel degree or place on distinct computing devices.

Single GPU Optimization

• sageattention for USP, we will for a stable version.
• More efficient Attention Implementations, remove dependencies on yunchang. support advanced attention implementations (FA3, FlashInfer, xformers, etc.) #319

Offloading

• Flux.1 parameter CPU offloading during inference

LoRA

• Fast LoRA Swapping (waiting for huggingface/diffusers@main...datacrunch-research:diffusers:dc-lora [WIP][LoRA] Implement hot-swapping of LoRA huggingface/diffusers#9453)
• CogVideoX-Fun-V1.1-Reward-LoRAs

[antferdom]

#324

[antferdom]

Model Zoo: add OmniGen