面向大模型推理的存算协同理论综述

摘要：大模型推理在长上下文、高并发与多轮复用条件下，受算力、存储容量与跨层数据移动共同约束。量化、缓存、卸载与阶段解耦常组合部署，但压缩率、命中率、显存占用下降等局部指标改善，并不等价于端到端关键路径缩短。从权重、键值（KV）缓存与激活的生命周期出发，以关键路径净收益为判据，系统梳理以算换存、以存换算与存算路径3类方法，并归纳其适用条件、附加开销、质量约束与失效边界，为面向负载的存算协同部署提供可复用的分析视角。认为未来应从生命周期控制、压缩表征、跨引擎接口与真实负载评测4个方面开展存算协同相关研究。

关键词：大模型推理；存算协同；数据移动；KV缓存；资源置换；瓶颈迁移

Abstract: Under long context, high concurrency, and multi-turn reuse, large model inference is constrained by computation, storage capacity, and cross-tier data movement. Quantization, caching, offloading, and stage disaggregation are often combined, but improvements in compression ratio, hit rate, or memory footprint do not imply a shorter end-to-end critical path. We start from the lifetimes of weights, key‑value (KV) caches, and activations, and using critical‑path net benefit as the evaluation criterion, this paper systematically reviews three categories of methods: computation for storage, storage for computation, and migration for capacity. It further summarizes their applicability, overheads, quality constraints, and failure modes, and provides a reusable analytical lens for workload‑oriented compute‑storage co‑optimization. This paper holds that future research related to compute‑storage co‑optimization should be conducted from four aspects: lifecycle control, compressed representation, cross‑engine interfaces, and real‑workload evaluation.

Keywords: large model inference; compute-storage co-optimization; data movement; KV cache; resource exchange; bottleneck migration