Worked on enhancing GPU observability in the NVIDIA/NeMo-Run repository by implementing a feature for collecting GPU device metrics using Nsys during Slurm-enabled runs. Developed logic to enable metrics collection conditionally on rank 0 through integration with the SlurmExecutor, ensuring efficient resource usage. Constructed the Nsys entrypoint to include metrics collection when applicable and added unit tests to verify correct behavior. Utilized Python and shell scripting to support distributed systems and performance profiling needs. This work improved the ability to analyze and optimize GPU workloads, providing better insight into system performance without introducing unnecessary overhead or complexity.