In the field of machine learning, distributed learning has gained significant attention as an effective approach for model training, enabling machines in different locations to collaborate in the training process without the need to transfer data to a central server. However, a key challenge is the limited computational capacity of local nodes, which restricts their ability to efficiently handle large-scale tasks. To address this issue, robust distributed learning frameworks are essential for optimizing performance. In this paper, we propose a novel approach that integrates the Multi-access Edge Computing (MEC) system with the Stackelberg game model to develop an optimal training task offloading scheme. This approach optimizes system’s computing resources while jointly meeting the delay requirements of training tasks. We will conduct extensive experiments to validate the effectiveness of this method against established benchmarks, including Non-offloading and ODO methods [11].