[BUG] LLava-RLHF run with BFloat16 failed

[hxhcreate]

In your implemetation at llava_rlhf_chat.py, you set dtype and compute type ad torch.BFloat16 which failed with

mat1 and mat2 must have the same dtype but got Half and BFloat16

change it to torch.float16 solve this problem

[zycheiheihei]

Thanks for pointing out. We have fixed this.