Aonier focused on enhancing instrumentation and observability for GPU memory usage during Megatron-LM training in the ROCm/Megatron-LM repository. They developed a feature that logs GPU memory utilization percentage throughout training, appending this data to the training log to provide better visibility into resource consumption. Using Python and leveraging deep learning and GPU computing expertise, Aonier implemented memory usage recording to support data-driven capacity planning and performance optimization for large-scale training runs. The work demonstrated a targeted approach to performance monitoring, addressing the need for actionable insights into GPU resource allocation without introducing unnecessary complexity or unrelated bug fixes.