面向大模型推理的存储优化研究综述OA

The trillion-parameter scale of large models places higher demands on storage perform-ance,and existing storage optimization solutions for large model inference face challenges such as insuf-ficient key-value cache capacity and excessive energy consumption during inference.This paper system-atically reviews optimization research in large model inference within the storage field in the past three years from three perspectives of hardware,software,and architecture.It summarizes the technical advantages of new hardware such as processing-in-memory(PIM)chip and optical interconnect,analyzes software optimization strategies including key-value cache management and data scheduling,and provides an overview of current storage architecture design methods for large model inference.During the review,the paper summarizes the challenges in storage-related research for large model inference,as well as the limitations of existing optimization techniques across different inference scenarios.At the end of this paper,future research directions are outlined in light of current trends in storage technology,providing guidance for storage optimization in large model inference.