M1T5 — Produce a GPU health report¶
Run:
The diagnostic is read-only. It writes results/health.json for diffing and results/health-report.md for review. It is deliberately useful on a broken or CPU-only host: missing nvidia-smi, a CPU-only PyTorch wheel, or an empty device mask becomes evidence instead of a Python traceback.
Success means the report names the driver, PyTorch build CUDA, visible devices, topology, active compute processes, clocks, power, ECC mode, and pending retired pages—or explicitly says why a field could not be read.