Worked on performance optimization for the yhyang201/sglang repository, focusing on the Aiter model’s attention mechanism. Addressed memory efficiency and throughput by removing the FP8 key-value upcast and implementing a native FP8 cache path, allowing the model to process larger sequences with reduced latency and resource consumption. The solution was developed in Python and leveraged deep learning and machine learning techniques, with the changes integrated through a collaborative pull request. This work improved inference efficiency by maintaining key-value caches in native FP8, setting a foundation for handling larger prompts while reinforcing code quality through peer review and collaborative development practices.