Worked on enhancing GPU backend capabilities in the tenstorrent/vllm and jeejeelee/vllm repositories, focusing on build systems and performance optimization. Developed a HIP-CUDA compilation interoperability feature that allowed HIP source files to be directly included in ROCm builds, streamlining workflows for developers working with both HIP and CUDA. Later, implemented a Sparse MLA performance enhancement that enabled MTP lens values greater than one in the Sparse MLA backend, increasing flexibility and throughput for larger workloads on ROCm. Leveraged CMake and Python to improve build compatibility, maintainability, and scalability, aligning code quality and testing with evolving ROCm performance goals.