Worked on the ml-explore/mlx-lm repository to deliver a feature focused on batch generation memory management and performance optimization. Addressed server responsiveness by introducing a wired memory usage limit, using a device-capability aware mechanism to dynamically set and reset this limit based on hardware constraints. Ensured that resources were properly closed to prevent memory leaks, improving reliability under concurrent workloads. The implementation was carried out in Python, leveraging back end development skills and a strong emphasis on memory management. This work enhanced the server’s ability to handle demanding tasks efficiently, resulting in better resource utilization and more stable performance overall.