Ash Bhandare developed GPU metrics instrumentation for the NVIDIA/NeMo-Run repository, focusing on enhancing observability and performance analysis for distributed GPU workloads. He implemented a feature that collects GPU device metrics using Nsys on rank 0 during Slurm-enabled runs, integrating this capability with the SlurmExecutor to conditionally enable metrics collection. The solution involved constructing the Nsys entrypoint to include metrics when appropriate and adding unit tests to ensure correct functionality. Leveraging Python, shell scripting, and system administration skills, Ash’s work addressed the need for improved resource optimization and monitoring, delivering a targeted, well-tested enhancement to the project’s profiling capabilities.