This is different from language models like GPT-2,
  which use autoregressive decoding instead of parallel.