End-Edge-Cloud Collaborative Inference for Efficient Large Language Model Services: A Survey

As large language model (LLM) services expand toward distributed environments characterized by latency sensitivity, heterogeneous resources, and stringent privacy constraints, neither end-only nor cloud-only deployment can simultaneously balance generation quality, response latency, resource cost, and energy consumption. This survey systematically reviews optimization techniques for end-edge-cloud collaborative LLM inference services and establishes a four-layer taxonomy encompassing the algorithm, model, system, and chip levels. This taxonomy captures the end-to-end execution chain from collaborative decision-making to physical execution. The algorithm layer determines inference paths and task partitioning; the model layer provides deployable and collaborative model forms; the system layer manages resource mapping and pipeline execution; and the chip layer supplies the underlying computing, storage, and interconnect capabilities. These four layers are mutually dependent and collectively determine the end-to-end benefits of a collaborative solution. Furthermore, this survey defines representative collaboration paradigms and evaluation metrics, compares the performance gains, resource overheads, and applicability boundaries of different methods, and summarizes promising future directions, including cross-layer collaboration, dynamic adaptation, state-centric runtimes, and service-aware hardware.