First run

Public self-service leads with H100-SXM. --gpu H100 maps to that SKU.

1. Add credit

compute credits add 10

Minimum top-up is $10. Complete Checkout in the browser, then:

compute credits

The balance updates after the payment webhook, not at the moment the Checkout tab opens. See Credits.

2. Write a function

Save as simple_mlp.py:

import compute

app = compute.App("simple-mlp")
image = compute.Image.cuda_pytorch()


@app.function(gpu="H100-80GB", image=image, timeout=600)
def train(steps: int = 40, hidden: int = 32, seed: int = 0) -> dict:
    import torch
    import torch.nn as nn

    torch.manual_seed(seed)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = nn.Sequential(
        nn.Linear(16, hidden),
        nn.ReLU(),
        nn.Linear(hidden, 2),
    ).to(device)
    opt = torch.optim.Adam(model.parameters(), lr=1e-3)
    loss_fn = nn.CrossEntropyLoss()
    x = torch.randn(256, 16, device=device)
    y = (x.sum(dim=1) > 0).long()
    last = None
    for _ in range(steps):
        opt.zero_grad()
        last = loss_fn(model(x), y)
        last.backward()
        opt.step()
    return {
        "ok": True,
        "compat": "mlp",
        "device": str(device),
        "cuda": bool(torch.cuda.is_available()),
        "steps": steps,
        "loss": float(last.detach()),
        "torch": torch.__version__,
    }

gpu="H100-80GB" is the same SKU as H100-SXM. Arguments and the return value must be JSON.

3. See what would ship

compute run simple_mlp.py::train --gpu H100-SXM --dry-run

The dry run prints the file set and manifest. Nothing is uploaded. No machine is created.

4. Run it

compute run simple_mlp.py::train --gpu H100-SXM --wait --yes --args '{"steps": 40, "seed": 20260814}'

--wait streams status (and logs when present) until a terminal state.
Without --wait, the CLI exits after create; use compute logs <run_id> -f later.
--yes skips the interactive price confirmation.

When it succeeds you get JSON back. Then:

compute runs receipt <run_id>
compute machines

compute machines should be empty after teardown. That is the product: one run, one machine, then gone.

If create is refused

A new run can be refused by the account’s one-active-run, spend, or create-rate limits, or by provider availability, even with a positive balance. compute doctor --json and the error body include a request id — send that to Support.

Overview

Get started

CLI

Reference

1. Add credit

2. Write a function

3. See what would ship

4. Run it

If create is refused

​1. Add credit

​2. Write a function

​3. See what would ship

​4. Run it

​If create is refused

1. Add credit

2. Write a function

3. See what would ship

4. Run it

If create is refused