Worked on performance optimization in the sgLANG attention path within the ping1jing2/sglang repository, focusing on improving metadata processing for attention mechanisms. Developed and delivered a fused Triton kernel for the normal_decode_set_metadata function, which reduced the number of sequential CUDA kernels required and streamlined the metadata path. This approach enabled higher throughput and lower latency in the attention pipeline, directly benefiting deep learning workloads. Collaborated effectively with a co-author throughout the development and code review process. Utilized Python, CUDA, and Triton to implement the solution, demonstrating depth in GPU programming and a targeted approach to machine learning system optimization.