Synergizing large and small models in cloud-edge continuum: a spatiotemporal hypergraph approach for dynamic offloading

Cloud-edge collaborative inference, particularly the synergy between edge-deployed small models and cloud-based large models, demonstrates extensive application value in meeting the high-accuracy and low-latency demands of emerging artificial intelligence. However, existing dynamic offloading methods exhibit limitations in modeling the “many-to-many” high-order resource competition that arises when numerous small models concurrently access shared large model interfaces. They lack a unified representation capable of adapting to time-varying network congestion and often ignore the spatial exclusivity of concurrent tasks when sharing bandwidth and large model interfaces. To overcome these limitations, this paper proposes a spatiotemporal hypergraph-driven framework integrating high-order topological feature extraction with dynamic resource modeling. This framework introduces dynamic hypergraph sequences to naturally encompass local conflict domains, mitigating the topological blind spots and “over-smoothing” issues inherent in traditional pairwise graphs. In the first stage, a spatiotemporal hypergraph neural network (ST-HGNN) encodes complex spatial dependencies and predicts temporal resource bottlenecks. Subsequently, a weighted multi-dimensional hypergraph matching (WTHM) strategy transforms the NP-hard mixed-integer nonlinear programming problem into an efficient heuristic search. Trace-driven simulation results based on real-world urban trajectories (EUA and T-Drive) demonstrate that the proposed framework achieves a robust normalized VCI score of 0.64 even during extreme congestion peaks. By consistently achieving an absolute VCI improvement of 0.17 to 0.38 over traditional bipartite-graph and proximity-based baselines, the framework significantly mitigates cascading queuing failures. Moreover, the developed dual-timescale asynchronous architecture strictly confines online execution to the millisecond scale, validating its capability to balance inference accuracy and latency across dynamic cloud-edge environments.