Worked on the intel/sycl-tla repository to enhance the correctness and reliability of the Mixture-of-Experts (MoE) inference and training path. Addressed a critical issue by fixing matrix dimension alignment in the MoE down-projection, ensuring that the grouped GEMM launcher receives accurate hidden and intermediate dimensions. This change prevents shape-related runtime errors and improves model stability across various MoE configurations. The work involved careful reasoning about tensor shapes and precise commit-level updates, utilizing C++, SYCL, and GEMM. The fix improved the robustness of the MoE implementation, reducing debugging time and post-deployment risk for complex model workflows.