Over a two-month period, contributed to the sglang repositories by developing advanced backend features focused on deep learning model inference and server optimization. Delivered NPU-accelerated DLLM LLaDA2.x graph mode inference for ping1jing2/sglang, optimizing model configurations and attention mechanisms to enhance throughput and reduce latency. In sgl-project/sglang, implemented dynamic page size configuration in the Scheduler, enabling configuration-driven scheduling and improved server argument management. Collaborated closely with cross-functional teams, integrating features such as the Dllm radix cache. Leveraged Python, deep learning, and NPU optimization skills to lay the groundwork for scalable, high-performance backend and inference workflows.