Worked on performance optimization for AMD ROCm tensor reductions in the pytorch/pytorch repository, focusing on three-dimensional tensor operations. Developed a feature that limits the number of values each thread processes during reductions, effectively reducing per-thread workload and improving throughput on AMD GPUs. The approach leveraged expertise in CUDA, GPU programming, and parallel computing, with implementation in C++. By capping per-thread workload, the solution addressed overhead issues inherent in large-scale tensor reductions, resulting in more efficient execution. The work demonstrated a targeted, technical solution to a specific performance bottleneck, contributing to improved performance for PyTorch users on AMD hardware.