arXiv · Computation and Language· Geonmo Gu, Byeongho Heo, HeeJae Jun, Yoohoon Kang, Sangmin Lee, Sangdoo Yun, Dongyoon Han·· 4 hr agoAI score33
DLoop:循环式投机解码
DLoop: Looped Speculative Decoding
AI brief
DLoop 是一种循环式投机解码方法,在草稿模型保持高置信度时连续执行多个草稿阶段,再统一验证累积的草稿 token,从而减少目标模型前向传播次数。
Source: arXiv · Computation and Language · arxiv.org