Is there a way to control training parameters like precision, batch size etc. to make this run on low-end gpus? Thanks