gpus = [0, 1, 2, 3] torch.cuda.set_device 'cuda:{}'.formatgpus[0] train_