Resource prediction in heterogeneous cloud environments is hard because of diverse node configurations, monitoring metrics, execution logs, and injected failures. Neural models can capture temporal patterns but miss sparse discrete features. Tree-based models can model categorical data but cannot handle complex spatio-temporal patterns. We propose MHST-GB, a Multi-Modal Hierarchical Spatio-Temporal Ensemble Network with Gradient Boosting Integration. The framework uses modality-specific neural encoders with correlation-guided attention for fusion. It combines a dual-path design of deep spatio-temporal networks and LightGBM for complementary feature spaces. It also adds a feedback-driven training method that adjusts attention weights based on feature importance. With curriculum learning, FMSE loss, Mixup, and DropConnect, MHST-GB gains robustness and generalization for accurate multi-resource prediction in heterogeneous cloud environments.
Large language models show strong abilities in code generation. They still face problems in semantic understanding, structural consistency, controllability, and generalization. Existing methods often focus on one aspect, like instruction following or structural accuracy. They do not offer a unified framework that balances correctness, readability, and adaptability. This work presents CodeFusion-Qwen72B, a multi-stage structure-aware tuning framework for Qwen-72B. It includes progressive low-rank adaptation, hybrid instruction optimization, multi-context fusion, structure-preserving hybrid loss, controllable generation decoding, and adaptive prompt retrieval. By improving semantic comprehension, structural alignment, and controlled code generation, the framework increases the robustness and versatility of large-scale models for software engineering tasks. This study shows that multi-stage optimization and structure-aware learning are effective ways to advance code generation with ultra-large models.