Worked on the ROCm/aiter repository to deliver strided q_nope tensor support within the fused QK RoPE cache kernel for MLA, addressing the challenge of handling non-contiguous memory layouts in GPU workloads. Updated the CUDA kernel to accurately compute memory offsets for these layouts, ensuring correct operation during both prefill and decode phases. Expanded test coverage to validate the new functionality and improve reliability in production scenarios. Collaborated across teams to co-author changes that strengthened MLA-path integration. The work leveraged C++, CUDA, and PyTorch, demonstrating depth in GPU programming and a focus on robust, production-ready kernel and test development.