During training, both BART and T5 will make the appropriate
  decoder_input_ids and decoder attention masks internally.