preemption-safe-training-template

A minimal PyTorch training loop that survives spot / interruptible GPU preemption. Copy train.py and swap in your model and data.

What it does

1. Keeps all recovery state (weights, optimizer, RNG, step) in one directory: $NODUS_STATE_DIR or ./state.
2. Writes checkpoints atomically (temp file, then os.replace).
3. Resumes from the last checkpoint on startup.
4. Saves on a timer, and again immediately when the platform sends a checkpoint request.

Run locally

pip install torch
python train.py          # Ctrl+C midway, run again: it resumes

Run on a cheap interruptible GPU with Nodus

Nodus copies the state directory off the machine before a reclaim and restores it on the replacement, so preemptions cost minutes instead of the whole run.

pip install nodus-compute
nodus login
nodus run --gpu L4 --image nodus/pytorch --interruptible --checkpoint /nodus/state -d -- python train.py
nodus logs -f job/<name>

New accounts get a $30 starter grant (valid for 30 days), so the first runs need no card. Add --dry-run to see the cost estimate before anything starts.

Picking a checkpoint interval

Aim for at least 4 checkpoints per expected run and keep time spent saving under about 10% of runtime. Set SAVE_EVERY_SEC accordingly.

More

• Full guide: https://www.nodus-compute.ai/docs/guides/checkpoints/
• Write-up: https://dev.to/viswaretas_kotra/how-to-make-gpu-training-survive-spot-preemption-without-babysitting-it-44k6
• GPU price index dataset: https://huggingface.co/datasets/Viswa-k/cloud-gpu-pricing