GPU detection

Vendor	Read from	Notes
NVIDIA	nvidia-smi	Whole GPUs and MIG slices with a known profile.
AMD	amd-smi / rocm-smi	Whole GPUs only. Partition modes (SPX/DPX/TPX/QPX/CPX, NPS) are recorded but are not isolation. Render nodes are mapped from PCI inventory.
Apple silicon	ioreg	GPU cores and unified memory. Jobs run natively (mlx, pytorch-mps, llama.cpp), not in Docker.

Each GPU reports model, memory, device class (whole, mig, vgpu, sriov, AMD partitions…), isolation, a hashed identity, and health (XID, ECC, RAS, exclusive mode). Only sensors the driver actually printed count; anything absent stays unknown.