Large language models have garnered significant attention and are widely utilized across different fields due to their impressive performance. However, centralized training of these models can pose privacy risks like data leakage, hindering their application and advancement. As a decentralized learning paradigm, Federated Learning(FL) can effectively address data security in large language models and has specific application scenarios. In addition, the fine-tuning process of large language models also incurs significant communication expenses and leads to the loss of computational resources. The utilization of prompt tuning in federated learning is anticipated to decrease training costs, accelerate the training process, and enhance model performance. In this paper, we propose ‘FedLLM-PPS’ to explore a novel approach for large language models within federated learning. This method address data security concerns associated with large language models while enhancing performance and efficiency. Experiments based on several datasets for various Natural Language Processing(NLP) tasks demonstrate superior performance and efficiency compared to baseline methods.