Worked on the huggingface/torchtitan repository to address a critical issue in Mixture of Experts (MoE) bias updates, focusing on improving both correctness and computational efficiency. Using Python and PyTorch, the developer fixed a double-counting problem during recomputation, ensuring that bias updates in MoE layers are applied accurately. The solution also optimized expert usage tracking, reducing unnecessary computations and enhancing training throughput. This targeted bug fix improved the reliability and reproducibility of large-scale deep learning experiments, enabling better resource utilization. The work demonstrated strong skills in algorithm optimization and deep learning, with careful attention to both performance and correctness.