Worked on enhancing observability for model training in the NVIDIA/Megatron-LM repository by developing advanced logging features. Introduced activation logging and tokens-per-expert logging to capture detailed forward-pass activation states and routing metadata, providing deeper insights into model behavior during large-scale training. Leveraged Python and PyTorch to implement these analytics, enabling more effective debugging and performance profiling. Collaborated closely with other contributors to ensure code quality and thorough review. This work laid the foundation for improved issue resolution and future optimizations in distributed training environments, reflecting a strong focus on deep learning, machine learning, and robust model training infrastructure.