Skyward

# pi.py
import skyward as sky


@sky.app(
    provider=sky.RunPod(),
    accelerator=sky.accelerators.RTX_3090(),
    image=sky.Image(pip=["torch", "numpy"]),
)
def estimate_pi(n: int = 100_000_000) -> dict[str, str | float]:
    """Estimate pi from N random points, on a GPU."""
    import torch

    points = torch.rand(2, n, device="cuda")
    inside = (points[0] ** 2 + points[1] ** 2 <= 1).sum().item()

    return {"gpu": torch.cuda.get_device_name(), "pi": 4 * inside / n}

sky server start
sky run -f pi.py --n 500000000

Installation

uv add "skyward[all]"

all includes every optional extra. For a smaller installation, combine only what you need, e.g. uv add "skyward[client,server,cli,aws]". See available extras.

Requirements

• Python 3.12+
• Cloud provider credentials (setup guide)

Documentation

Full documentation at gabfssilva.github.io/skyward.

License

MIT