DSpark: Speculative decoding accelerates LLM inference

DeepSeek's DSpark paper presents a speculative decoding method for faster LLM inference. The technique accelerates generation by predicting multiple tokens in parallel.