DSpark: Speculative decoding accelerates LLM inference [pdf]github.com
559 points by aurenvale 6 hours ago | 217 comments