ysystudio

seem like the model after training using torch.cuda.amp's autocast(), its dtype is fp32, if want to deploy the model , dose it need covert to fp16 manualy? it is little...