5fa1a76
1
2
This is different from language models like GPT-2, which use autoregressive decoding instead of parallel.
