Worked on optimizing the attention mechanism in the ROCm/aiter repository by developing experimental pa_ragged kernels aimed at improving deep learning model throughput. The approach involved implementing double-buffered K-cache loading and non-temporal key-value loads to enhance memory access patterns. Leveraging GPU programming techniques, a 64-thread path was created to efficiently load the K-cache into local data storage and distribute data in alignment with MFMA requirements. The work was carried out using C++ and CUDA, with added unit tests to ensure reliability. This feature-focused contribution demonstrated depth in performance optimization and advanced memory management for GPU-accelerated deep learning workloads.