use torch_dtype=fp16 when specified

[tmm1]

i noticed this while reviewing the code. when using --fp16, it still sets torch_dtype=torch.float32

is this by design, or an oversight?

is this intentionally set to a different value than compute_dtype?

[tongyx361]

I think it should be a typo error, because there is another torch.float32 at the end of if-else logic.