Questions about training vision encoder

Hi, guys.

Eagle is a really nice work, and models work very well in practice. I just have some questions about vision encoder's training details.

Do you guys train all vision encoders or keep some fixed? Why decorate the forward function of CLIP with `no_grad` [here](https://github.com/NVlabs/EAGLE/blob/main/eagle/model/multimodal_encoder/clip_encoder.py#L46)?