Skip to content
Nodus

Run your own code

View Markdown

This page runs a script from your own directory on a GPU, then downloads the file it wrote. It assumes you have installed the CLI and signed in. Every command here runs in CI as the cli/quickstart example.

  1. Put your code in a directory. Any directory works; this one holds a small PyTorch training loop that saves its final metrics to /nodus/outputs/metrics.json.

    train.py
    """Fit a small linear model, print progress and save metrics.json as the declared output "metrics"."""
    

    import json
    import pathlib
    

    import torch
    

    device = "cuda" if torch.cuda.is_available() else "cpu"
    print("device:", torch.cuda.get_device_name() if device == "cuda" else "cpu")
    

    torch.manual_seed(0)
    x = torch.randn(1024, 16, device=device)
    y = x @ torch.randn(16, 1, device=device)
    w = torch.zeros(16, 1, device=device, requires_grad=True)
    opt = torch.optim.SGD([w], lr=0.1)
    

    for step in range(200):
        loss = ((x @ w - y) ** 2).mean()
        opt.zero_grad()
        loss.backward()
        opt.step()
        if step % 50 == 0:
            print(f"step {step} loss {loss.item():.4f}")
    

    print(f"final loss {loss.item():.6f}")
    out = pathlib.Path("/nodus/outputs")
    out.mkdir(parents=True, exist_ok=True)
    (out / "metrics.json").write_text(json.dumps({"final_loss": loss.item(), "steps": 200}))

  2. Run it. From that directory:

    Terminal window
    nodus run --name cli-quickstart --gpu L4 --image nodus/pytorch \
      --output metrics=/nodus/outputs/metrics.json -- python train.py
    Terminal window
    job/cli-quickstart created · est. $0.01–0.03 · starts in ~2–4 min (cold) · hold $0.50 · Ctrl+C to cancel, -d to detach
     ✓ Scheduled      l4-24g-x1-us · $0.52/h (rate frozen)
     ✓ Provisioning   1m48s
     ▶ Running
    device: NVIDIA L4
    step 0 loss 15.8732
    ...
    final loss 0.000001
     ✓ Succeeded in 2m31s · $0.02 (boot $0.01 · running $0.01) · kept warm 60 s · nodus describe job/cli-quickstart

    Before anything is charged you see the estimate and a hold: the most this run can cost before you are asked. nodus run then uploads the directory (files listed in .gitignore or .nodusignore stay local), prints each phase, streams your program’s output on stdout and ends with the cost line. The CLI exits with your command’s exit code, so it works in scripts and CI.

  3. Get the output. The --output flag declared metrics; copy it back:

    Terminal window
    nodus cp job/cli-quickstart:outputs/metrics metrics.json

    The file is checked against its SHA-256 digest and only written when it matches.

Leave it running

Section titled “Leave it running”

Press Ctrl+C during a run and the CLI asks whether to cancel the Job; answer n to detach and leave it running. Or start with -d to return as soon as the Job exists. A detached Job keeps running:

Terminal window
nodus get jobs                        # every Job in the project, with phase and cost so far
nodus get jobs --mine -w              # only yours, updating live
nodus logs -f job/cli-quickstart      # stream the output again
nodus describe job/cli-quickstart     # attempts, placement, cost, conditions and events
nodus cancel job/cli-quickstart       # stop it; you pay only for what ran

Control the cost

Section titled “Control the cost”
Flag Effect
--max-cost 5 The Job stops gracefully before it spends more than $5
--timeout 2h The Job is stopped after two hours of wall-clock time
--dry-run Print the estimate and exit without creating anything

nodus billing shows your balance and this month’s spend; nodus billing usage --group-by day breaks it down.

Clean up

Section titled “Clean up”

Finished Jobs are kept for 30 days so you can read their logs and outputs, then deleted. To delete one now:

Terminal window
nodus delete job/cli-quickstart

Next steps

Section titled “Next steps”