Worked on optimizing MXFP4 tensor operations within the llama.cpp repository, focusing on enhancing the performance of OpenCL kernels for GPU-accelerated inference. The approach involved kernel-level enhancements, function flattening, and improved memory management, all implemented using C++ and OpenCL. These changes led to measurable improvements in runtime and throughput for MXFP4 paths on supported GPUs, reducing latency and increasing efficiency. The work also included code quality improvements in the OpenCL backend, laying the groundwork for future optimizations and easier maintenance. Emphasis was placed on performance tuning and GPU programming to address the specific needs of high-throughput tensor operations.