The recommendation of medication for discharge is a key task in clinical decision support and is of great significance for improving medical quality. Although large language models perform well in natural language processing, they still face challenges in the medical field where data is scarce and professional barriers are high. This paper proposes a model discharge medication recommendation framework called DP-EMR that integrates data augmentation, efficient fine-tuning, and innovative inference: It expands the training data by applying a pseudo-label augmentation technique based on confidence screening to the validation set, performing efficient fine-tuning of multiple heterogeneous large language models using LoRA, designing an innovative two-stage hierarchical integration strategy. In this strategy, different models trained on the original data and enhanced data are first weighted and voted for integration respectively, and then the prediction results of the first stage were fused through an adaptive strategy based on the prediction length. This method significantly enhanced the accuracy and robustness of medication recommendation, and ultimately rank second in the CHIP2025 discharge medication recommendation Task.