DSpark: Speculative decoding accelerates LLM inference [pdf]