dont send dummy batch when reloading from checkpoint

            if not p.requires_grad:

                continue

            if p.grad is None:

                raise RuntimeError('Model parameter did not receive gradient: ' + name + '. '

                                   'Use the param in the forward pass or set requires_grad=False')

                print('WARNING: model parameter did not receive gradient: ' + name + '. '

                      'Check that you\'re using the param in the forward pass or set requires_grad=False')

                grads.append(p.new_zeros(p.shape))

            else:

                grads.append(p.grad.data)

        return grads

    )

    # Load the latest checkpoint if one is available

    load_checkpoint(args, trainer, epoch_itr)

    if not load_checkpoint(args, trainer, epoch_itr):

        # Send a dummy batch to warm the caching allocator

        dummy_batch = task.dataset('train').get_dummy_batch(args.max_tokens, max_positions)

        trainer.dummy_train_step(dummy_batch)

            trainer.lr_step_update(trainer.get_num_updates())

            if 'best' in extra_state:

                save_checkpoint.best = extra_state['best']

        return True

    return False

def load_dataset_splits(task, splits):