Allow torch tensor memory to be released and resumed later - View it on GitHub