Worked on backend performance improvements for the bytedance-iaas/sglang repository, focusing on optimizing the Mamba attention mechanism. Delivered a targeted refactor that removed unnecessary device-to-host (D2H) operations during state tracking, which enhanced both efficiency and code clarity in the attention processing path. The approach emphasized maintainability and throughput, supporting future scalability for deep learning workloads. Utilized Python and applied expertise in data structures and machine learning to streamline backend operations. Collaborated with other contributors through code reviews and co-authoring, demonstrating a methodical approach to backend refactoring and performance optimization without introducing new features or addressing bug fixes.