During January 2025, Aonier focused on enhancing instrumentation and observability for GPU memory usage within the ROCm/Megatron-LM repository. They developed a feature that logs GPU memory utilization during deep learning training, calculating usage percentages and appending this data to training logs. This approach, implemented in Python and leveraging GPU computing and performance monitoring skills, provided actionable insights for capacity planning and resource optimization in large-scale training environments. While the work was limited to a single feature and did not include bug fixes, it demonstrated depth in addressing performance visibility, enabling more data-driven decisions for managing computational resources during model training.