SAEM: Stage-Aware Expert Management for Memory-Efficient MoE Inference in Chain-of-Thought Reasoning
SAEM:面向思维链推理的内存高效混合专家模型推理的阶段感知型专家管理
机构 * School of Computing, National University of Singapore(新加坡国立大学计算机学院)
专题命中 数学推理 :reasoning(title,abstract);chain-of-thought(title,abstract);CoT(abstract,abstract_cn);分类 cs.AI
AI总结 针对思维链推理中MoE模型的内存与延迟问题,提出SAEM阶段感知型专家管理运行时,通过阶段感知缓存等技术提升吞吐量,在受限GPU内存下平均实现1.33倍的性能提升。
Comments Extended version of the paper accepted at DAC 2026. Extends the conference version with evaluations on AIME 2024 and GPQA-Diamond and a prediction-oracle upper-bound analysis