Discharge medication recommendation is a critical task in clinical decision support, aiming to assist physicians in determining safe and personalized prescriptions based on inpatient electronic health records (EHRs). However, the complexity of multi-disease comorbidities, long-tail medication distributions, and the scarcity of annotated data pose significant challenges for existing methods. In this paper, we propose a framework that integrates multi-dimensional feature enhancement and multi-scale model training to improve model accuracy and generalization. At the data level, we construct a feature-augmented EHR representation by incorporating pharmacological categories, patient meta-features, and disease–drug statistical priors, coupled with structural perturbation strategies to alleviate data sparsity. At the model level, we adopt a multi-scale fine-tuning strategy using Qwen series language models of different parameter sizes, followed by a two-level hierarchical ensemble that fuses complementary knowledge across model scales and augmentation variants. Experiments on the CHIP 2025 Shared Task 2 (Discharge Medication Recommendation for Metabolic Diseases) dataset demonstrate that our approach achieves state-of-the-art performance, surpassing both the official baselines and all other participating systems.