Continuum-Aware End-Cloud Offloading for Collaborative Inference

Distributed computing continuum systems (DCCSs) integrate heterogeneous end, edge, and cloud resources to support adaptive task execution across tiers. In such environments, collaborative inference with large and small language models offers a promising way to balance inference quality and execution efficiency. However, most existing collaboration strategies are model-centric and often overlook dynamic system conditions such as cloud congestion and device-side load. This paper presents a continuum-aware collaborative offloading (CACO) framework for end-cloud collaborative inference with edge-side scheduling and congestion sensing. The framework follows a cloud-first policy under uncongested conditions to preserve accuracy. When cloud congestion is detected, the edge gateway decides whether to activate end-side collaborative execution according to end-node load conditions and subjectspecific reliability priors. Reliable end nodes are selected for top- $k$ collaborative inference, and majority voting is used to generate the final answer. If consensus fails, the system falls back to cloud inference and updates the reliability priors online. We implement a prototype and evaluate it on the MMLU benchmark under different congestion levels. Experimental results show that the proposed framework preserves near-cloud answer quality under regular conditions while reducing cloud workload pressure, and achieves stronger latency robustness than static offloading under higher congestion.