Support for NVIDIA Cosmos 3 model family?

### Feature Idea

https://huggingface.co/collections/nvidia/cosmos3

4B (upcoming), 16B, 64B parameters

All support text, image, video, audio: both directions, inputs and outputs.

The post trained versions are SOTA open weights in their tasks (T2I, I2V), the T2I 64B one comes at number 4 on artificial analysis close to nano banana 2 showing that the community can easily get SOTA levels with SFT on the tasks they want which suggests that the base models are extremely adaptable and has a good arch.