Developed XPU multimodal support for the ERNIE-4.5-VL-28B-A3B model within the PaddlePaddle/ERNIE repository, enabling large-scale vision-language processing on XPU hardware. Focused on updating data handling and device compatibility, the work improved both performance and deployment scalability for multimodal tasks. Leveraged deep learning and distributed computing techniques, with model training workflows implemented in Python to ensure robust integration with existing infrastructure. The feature broadened enterprise use cases by supporting advanced multimodal models, addressing the need for efficient processing across diverse hardware environments. This contribution demonstrated depth in machine learning engineering and an understanding of scalable, production-ready model deployment.