About torch.cuda.empty_cache()

it is related to this question Pytorch abnormal inference time