Worked on enhancing Mixture of Experts (MoE) support and performance in the llama.cpp and ggml repositories, focusing on SYCL-backed components and matrix multiplication strategies. Developed robust handling for reordered MoE weights, including Q4_K, Q5_K, and Q6_K expert tensors, and introduced a new matrix multiplication method that optimizes data layout for improved throughput. Implemented graceful fallback mechanisms for unsupported edge cases, reducing runtime errors and broadening backend compatibility. Leveraged C++, SYCL, and GPU programming to deliver more reliable MoE inference, enabling smoother production deployments and laying the foundation for scalable, high-throughput machine learning workloads across supported backends.