Worked on optimizing CUDA argmax reduction algorithms in both the ggml and llama.cpp repositories, focusing on improving GPU throughput and inference speed for large language model workloads. Refactored the reduction offset logic to use WARP_SIZE/2, replacing hardcoded values to achieve a better balance between performance and accuracy in parallel reductions. Applied a consistent optimization pattern across codebases, aligning with upstream goals and enabling unified performance improvements. Leveraged expertise in CUDA, GPU programming, and parallel computing to enhance the adaptability and efficiency of the argmax path, resulting in faster model inference and improved GPU utilization without introducing new bugs.