Accurate prediction of wireless traffic is crucial for optimizing network resources and enhancing service quality. We explores the application of a modified Generative Pretrained Transformer architecture, mGPT, for wireless traffic prediction, leveraging its autoregressive Transformer architecture. A core feature of the model, the self-attention mechanism, enables it to directly capture relationships between any two points in a sequence, effectively modeling complex nonlinear dependencies and improving long-term predicting accuracy. By employing a causal attention mechanism, the model ensures that the predictions are based solely on past observations, resulting in reliable future forecasts. Data from public datasets are collected and preprocessed using a sliding window technique to convert raw traffic data into temporally rich sequence datasets, providing a solid foundation for analysis and prediction. Key structural modifications to the original GPT model include replacing its final output layer, originally designed for text classification, with a linear regression layer tailored for continuous value prediction. This adjustment improves the adaptability of the model for wireless traffic forecasting. The experimental results demonstrate that our proposed model significantly outperforms existing approaches in test sets.