Quick Start

Get up and running with Cumulus in minutes.

Installation

pip install cumulus-sdk

Initialize Client

from cumulus import CumulusClient

client = CumulusClient(api_key="YOUR_API_KEY")

Upload a Model

Register your local model with Cumulus:

endpoint = client.upload_model(
    name="detector_model",
    path="./models/detector.pth"
)

Deploy a Model

endpoint = client.deploy(
    model="detector_model",
    workload_type="inference"
)

# Get your endpoint ID
print(endpoint.id)  # Returns: "ep_1a2b3c4d5e6f7g8h"

Run Inference

result = endpoint("input_data_here")
print(result.output)

That's it. Cumulus automatically:

• Uploads and registers your model
• Analyzes your model size
• Finds the best GPU in the closest region
• Handles fractional allocation if needed
• Routes requests optimally

No configuration required. Just upload, deploy, and run.

Container Configuration & Dependencies

Cumulus handles container setup automatically, but you can specify custom dependencies:

endpoint = client.deploy(
    model="detector_model",
    workload_type="inference",
    dependencies=[
        "torch==2.4.0",
        "transformers==4.40.0",
        "accelerate==0.30.0"
    ],
    env_vars={
        "HF_HUB_ENABLE_HF_TRANSFER": "1"  # faster model downloads
    }
)

Cumulus automatically:

• Installs dependencies in the container
• Configures environment variables
• Optimizes the base image for your model type

Common Dependencies

Package	Description
torch	PyTorch for deep learning
transformers	Hugging Face model hub
vllm	High-performance LLM serving
accelerate	Distributed inference