ncu-report-skill

Run `npx skills add mit-han-lab/ncu-report-skill --all -g -y` in your terminal to install the skill globally.

The skill walks you through building a standalone harness, running ncu full and source-level profiles, and parsing results with the ncu_report Python API. It then guides analysis across six dimensions like occupancy, stalls, and memory.

The included diagnosis playbook maps Nsight Compute signals to likely causes and concrete fixes, such as load imbalance on variable-length inputs or L1 latency-bound stalls. It helps you cite exact metric values instead of guessing.

Yes. The workflow and helper scripts run on any GPU Nsight Compute supports, though metric names are optimized for B200 sm_100 and may differ on A100 or H100.

Yes. The Python helpers for metric extraction, stall hotspot ranking, and timeline plotting run standalone on any .ncu-rep file from the command line.