Excited to share our work with friends from MIT/Google on Learned Asynchronous Decoding! LLM responses often contain chunks of tokens that are semantically independent. What if we can train LLMs to identify such chunks and decode them in parallel, thereby speeding up inference? 1/N