<p>Efficient task offloading in the Internet of Vehicles (IoV) is crucial to meet the increasing demand for computational resources. However, the dynamic and heterogeneous nature of IoV networks presents significant challenges in selecting optimal offloading destinations across edge, cloud, or neighboring vehicles. To address these challenges, this paper proposes an innovative approach that integrates virtual network embedding (VNE) with graph attention networks (GAT) and hierarchical deep reinforcement learning (HDRL) to enhance task offloading performance in IoV environments. GAT enables the model to capture complex topological dependencies between IoV nodes, improving the accuracy of resource selection by considering both network dynamics and task requirements. HDRL facilitates sequential decision-making within a hierarchical framework, optimizing task offloading decisions at multiple levels (vehicle, edge, cloud) based on long-term system performance metrics such as latency, energy efficiency, and task acceptance rates. Our VNE framework intelligently maps virtual networks, designed to reflect task requirements, to the best available resources. This approach maximizes resource efficiency while minimizing offloading delays. Simulation results demonstrate that the proposed approach significantly improves task offloading in the IoV compared to traditional methods. It achieves lower latency, better resource utilization, and higher task acceptance rates, showcasing its potential to enhance the overall efficiency of task management in next-generation vehicular networks.</p>