Worked on performance optimization for the deepseek-ai/DeepEP repository, focusing on the Notify Dispatch: Metadata Calculation path. Implemented a CUDA-based solution in C++ that dynamically adjusts warp sizing to match the number of channels, allowing a single loop to process metadata for all channels efficiently. This approach reduced loop iterations and improved GPU throughput by aligning warp configuration with channel count, enhancing scalability and lowering latency in metadata preparation. The work demonstrated expertise in CUDA programming and performance optimization, resulting in more efficient resource utilization and maintainable code that supports easier future extensions without introducing any bug fixes during the period.