近年来，基于Transformer架构的大型语言模型（LLM）在文本生成、摘要和翻译等多种任务中表现出色，但由于其自回归特性，即逐个生成Token的特点导致推理速度慢且计算成本高。推测解码（Speculative Decoding）通过一个较小的辅助模型生成候选Token，然后由主模型进行验证，从而显著缩短LLM的推理时间。推测解码能够保持与多项式采样 (Multinomial ...

Editor’s note: In this year’s back-to-school season, many freshmen who were born in the 21st century, known as Generation Z, will start their college life.