Worked on the NVIDIA-NeMo/Megatron-Bridge repository to enhance the reliability of large-scale deep learning training by addressing a critical issue in learning rate scheduling. Focused on stabilizing and validating the learning rate warmup calculation, the developer corrected the logic to use total decay iterations multiplied by the global batch size, ensuring consistent behavior across training runs. Updated unit tests and configuration management scripts in Python to accurately reflect the new scheduling logic, which improved training stability and convergence. Leveraged skills in deep learning, machine learning, and testing to reduce the risk of mis-scheduled learning rates that could impact experimental outcomes.