Worked on the apple/axlearn repository to deliver a flexible exponential moving average (EMA) schedule enhancement for machine learning training workflows. Introduced a step_offset parameter to the ema_schedule, allowing adjustment of the initial training step for warm-up and enabling more adaptive EMA decay schedules. This Python-based solution improved training stability and reduced the need for manual hyperparameter tuning, supporting more reproducible experiments across diverse workloads. The implementation demonstrated careful API design and ensured backward compatibility with existing training flows, enhancing maintainability and extensibility. Emphasized robust testing practices to validate the new feature and maintain reliability within the codebase.