Is it possible to cache model objects trained on GPU via joblib? Saving/Loading model from disk is a hassle.