Worked on the intel/sycl-tla repository to stabilize GEMM kernel execution for SM90 architectures, focusing on resolving hangs and stream-K launch errors when beta equals one. Addressed these issues by refining synchronization logic, including precise placement of load_order_barrier instructions and synchronization points, and introduced a bypass parameter to handle edge cases in occupancy calculations. These C++ and CUDA-based changes improved runtime stability and predictability for high-performance computing workloads, reducing debugging cycles and supporting smoother deployment. The work demonstrated a strong grasp of low-level optimization and kernel development, enhancing the reliability of GEMM operations on advanced hardware within a short timeframe.