Python SDK Reference

The official novapod Python package provides typed, asynchronous, and synchronous APIs for orchestrating GPU compute directly from training pipelines, automated evaluation harnesses, or autonomous agent loops.

---

Installation

pip install novapod

---

Authentication

Set your API token via environment variable or pass it directly to the Client:

export NOVAPOD_API_KEY="np_live_your_secret_key"

---

Basic Usage

Context Manager (Auto-Cleanup)

from novapod import Client

with Client() as client:
    # 1. Search for cheapest available H100
    offers = client.market.search(gpu="H100_SXM", min_reliability=0.98)
    best_offer = offers[0]
    print(f"Deploying on {best_offer.id} at ${best_offer.price_per_hour}/hr")

    # 2. Provision instance
    instance = client.instances.create(
        offer_id=best_offer.id,
        template="pytorch-2.4",
        disk_gb=100,
        name="nightly-eval-run"
    )

    # 3. Block until instance is running (<15s)
    instance.wait_for_running()
    print(f"SSH command: {instance.ssh_command}")
    print(f"JupyterLab: {instance.jupyter_url}")

    # 4. Execute remote command over SSH
    output = instance.exec("nvidia-smi --query-gpu=name,memory.total --format=csv")
    print(output)

    # 5. Pause or destroy
    instance.pause()

---

Asynchronous SDK (AsyncClient)

For high-concurrency applications, FastAPI backends, and autonomous agent loops:

import asyncio
from novapod import AsyncClient

async def run_batch():
    async with AsyncClient() as client:
        instances = await asyncio.gather(*[
            client.instances.create(gpu_type="RTX_4090", template="vllm")
            for _ in range(4)
        ])
        print(f"Provisioned {len(instances)} nodes concurrently.")

asyncio.run(run_batch())