CLI: probe
GPU discovery and real-time telemetry. Enumerates GPU devices via the hwledger-probe crate (NVIDIA → AMD → Metal → Intel). JSON output is stable-schema (hwledger.v1) and suitable for downstream fleet ingestion.
What you'll see
Running hwledger probe displays:
- Each GPU's index, model name, VRAM (total and free)
- Compute capability / architecture (e.g. CUDA compute 8.9 for RTX 4090)
- Current memory utilization and temperature
- Driver version (NVIDIA) or ROCm version (AMD)
Watch as the probe detects your hardware in real-time. On multi-GPU boxes, you'll see each device listed separately.
Journey not yet recorded.
Run the journey recorder to capture interactions:
./apps/macos/HwLedgerUITests/scripts/run-journeys.shWhat to watch for
- GPU order: Index 0, 1, 2... maps to
CUDA_VISIBLE_DEVICES - Memory free: Critical for planning (if 0 GB free, model won't fit)
- Compute capability: Determines supported attention variants (older GPUs may not support newest optimizations)
- Temperature: Shows if GPU is thermally constrained
- Multi-GPU layout: Bandwidth between GPUs affects tensor parallelism efficiency
Next steps
- Plan for your model — use probe output to guide planning
- Fleet probe command — all flags and output formats
- Watch mode — continuous GPU monitoring
Reproduce
bash
# List GPUs (human-readable)
hwledger probe
# List as JSON (for scripting)
hwledger probe --json | jq '.gpus[] | {name, vram_free_gb}'
# Watch continuously
hwledger probe --watch