GPU count
rocm-smi reports all 8 compute dies?
30s via srun
ECC errors
Uncorrectable memory errors → corrupt tensors
Abort if found
Network (Slingshot)
NIC interface up? Down NIC → RCCL timeouts
Avoids 20min hang
NVMe writable
Local SSD full from previous user?
Ckpt depends on it
32-node health check simulation
Click "Run" to simulate
Run check
Reset