fp16 and fp32 doesn't change the generation time

#3
by troilus - opened

both torch.float16 and torch.float32 having the same generation speed (~7s per 20 steps in GPU)

Sign up or log in to comment