Developed enhanced instrumentation for GPU memory observability within the ROCm/Megatron-LM repository, focusing on supporting capacity planning and performance optimization during deep learning training. Implemented a feature in Python that logs GPU memory usage by calculating utilization percentages and appending this data to the training log, providing actionable insights into resource consumption. Leveraged expertise in GPU computing and performance monitoring to enable data-driven decisions for large-scale model training. The work emphasized improving transparency around memory usage, facilitating more accurate budgeting and resource allocation. No major bug fixes were recorded during this period, with efforts concentrated on feature development and monitoring improvements.