skyward 0.1.0

Execute Python functions on cloud compute with a simple decorator

Cloud accelerators with a single decorator

---

Skyward is a Python library for ephemeral accelerator compute. Spin up cloud accelerators, run your code, and tear them down automatically. No infrastructure to manage, no idle costs.

Quick Example

import skyward as sky

@sky.compute
def train(epochs: int) -> dict:
    import torch
    model = torch.nn.Linear(100, 10).cuda()
    optimizer = torch.optim.Adam(model.parameters())

    for epoch in range(epochs):
        loss = model(torch.randn(32, 100, device="cuda")).sum()
        loss.backward()
        optimizer.step()

    return {"final_loss": loss.item()}

with sky.ComputePool(provider=sky.AWS(), accelerator="T4", image=sky.Image(pip=["torch"])) as pool:
    result = train(epochs=100) >> pool
    print(result)

Features

• One decorator, any cloud — @compute makes any function remotely executable. AWS, RunPod, VastAI, and Verda with a unified API.
• Operators, not boilerplate — >> executes on one node, @ broadcasts to all, & runs in parallel. No job configs, no YAML.
• Ephemeral by default — Instances provision on demand and terminate automatically. Context managers guarantee cleanup.
• Multi-provider support — AWS, RunPod, VastAI, Verda with automatic fallback and cost optimization.
• Distributed training — PyTorch DDP, Keras 3, JAX, TensorFlow, and HuggingFace integration decorators.
• Distributed collections — Dict, set, counter, queue, barrier, and lock replicated across the cluster.
• Spot-aware — Automatic spot instance selection, preemption detection, and replacement. Save 60-90% on compute costs.

Install

uv add skyward

Requirements

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

Documentation

Full documentation at gabfssilva.github.io/skyward.

License

MIT