Contributed to deep learning infrastructure by enhancing deployment stability and model flexibility across two major repositories. In paddlepaddle/paddleocr, addressed environment compatibility by updating the OpenCV version in the Dockerfile, which reduced deployment failures and improved reliability for containerized workflows using Docker and DevOps practices. In huggingface/diffusers, implemented native NPU attention mask shape support, enabling models to handle a broader range of inference scenarios on NPU hardware. This feature, developed in Python and PyTorch, improved the adaptability of attention mechanisms and reduced the need for workarounds. Work demonstrated attention to maintainability, code quality, and end-to-end traceability through clear commit history.