Avoid the problem of multiple nodes loading each time when the model is the same and the problem of GPU memory release.