Worked on improving the reliability of distributed training in the ml-explore/mlx repository by addressing a critical issue in the bias handling path of the all_to_sharded function. Using Python and leveraging expertise in distributed systems and machine learning, implemented a targeted fix that corrected the bias shard axis, ensuring accurate representation of bias weights in sharded configurations. This adjustment reduced the risk of training instability and improved reproducibility across distributed setups. The work also included validating and cleaning up the bias handling logic, laying the groundwork for future scalability and enhancing the overall stability of distributed sharding workflows within the project.