Support for WAN2.2 FP8 Quantization + Sage Attention + LightX2V LoRA Inference

Hello,

I am using your library to run WAN2.2 image-to-video generation, and I have encountered some issues.

Feature request / question:
Could you please add support for WAN2.2 FP8 quantization + Sage Attention + LightX2V LoRA inference, similar to what Kijia’s ComfyUI workflow provides?
From my understanding, this is currently one of the best approaches to maintain model generation quality while achieving the fastest inference speed.
If this integration is too complex to implement directly, could you please suggest where in the codebase I should start modifying things to implement this myself?

Second question:
Is it possible to run two 14B MoE models in a streaming/rotation setup, to avoid the overhead of loading and unloading the models each time?
Any suggestions or best practices on how to achieve this would be greatly appreciated.

Thank you very much for your work!

Start by locating the WAN2.2 inference entry point and any existing FP8 quantization, Sage Attention, or LightX2V LoRA integration. Then investigate whether the model-loading path supports keeping two 14B MoE models resident for streaming or rotation. Done means the requested inference workflow is supported and the model-loading behavior is documented or implemented.