Skip to main content

Quickstart

Sign up, install the CLI, and run your first GPU job in under five minutes.

1. Create an account​

Go to account.yarn.prosodylabs.com.au and sign up. Every new account gets $5 free compute balance to start.

2. Install the CLI​

pip install yarn-au

3. Authenticate​

yarn auth login

This opens your browser for login. Once authenticated, your credentials are saved locally.

4. Check available GPUs​

yarn gpus

You'll see available GPUs with pricing and current availability. Platform GPUs (Perth, AU) are free during beta.

GPUVRAM$/hr
RTX 409024 GB$1.50
A100 40GB40 GB$3.50
A100 80GB80 GB$5.00
H10080 GB$8.00

5. Submit your first job​

Create a file called train.py:

import yarn.train as yt
import torch
import torch.nn as nn

@yt.model
def net():
    return nn.Sequential(nn.Linear(784, 256), nn.ReLU(), nn.Linear(256, 10))

@yt.dataset(batch_size=64)
def mnist():
    from torchvision import datasets, transforms
    t = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
    return {"train": datasets.MNIST("./data", train=True, download=True, transform=t)}

@yt.job(epochs=5)
def train(model, data):
    opt = torch.optim.Adam(model.parameters(), lr=1e-3)
    for epoch in range(5):
        for x, y in data["train"]:
            loss = nn.functional.cross_entropy(model(x.view(-1, 784).cuda()), y.cuda())
            opt.zero_grad(); loss.backward(); opt.step()
        yt.report(loss=loss.item(), epoch=epoch)

train()

Check if it fits before submitting:

yarn job submit train.py --dry-run

This prints parameter count, memory breakdown, GPU fit verdict, and estimated cost -- without actually submitting.

Submit it:

yarn job submit train.py --gpu rtx-4090

Watch the logs:

yarn job logs <job-id> --follow

What next?​