During June 2026, contributed to the pytorch/torchtitan repository by addressing a targeted bug in the Mixture of Experts (MoE) load-balancing mechanism. The work focused on ensuring that only enabled layers update their expert bias and token counts, which corrected a regression affecting mixed-enabled configurations. This solution, implemented in Python using PyTorch, improved both the correctness and efficiency of the load-balancing process. Validation was performed through comprehensive unit testing with pytest, confirming robust behavior under various scenarios. The fix reduced unnecessary computations during training and enhanced the scalability and stability of large-scale MoE models within the repository.