Does this model support float16 or bfloat16 when running on CPU?

[luentong]

I compiled float16 and bfloat16 calculation library for my cpu, but when I set dtype=torch.float16/bfloat16, weights get converted back to float32 during inference.
Thanks