Skip to main content

Cumulus

Run GPU workloads without managing infrastructure. Cumulus provides GPU compute on NVIDIA A100/H100 GPUs, so you only pay for what you use.

What is Cumulus?​

Cumulus is a GPU compute platform for ML workloads. Submit your training scripts and Cumulus handles the infrastructure - scheduling, scaling, and GPU management.

Key benefits:

  • Flexible GPU allocation - Request the resources you need, from partial to multiple GPUs
  • Automatic checkpointing - Training jobs save progress and resume if interrupted
  • Zero infrastructure - No Kubernetes, Docker, or GPU drivers to manage
  • Pay per use - Only pay for GPU time you actually consume

Quick Start​

pip install cumulus-sdk
from cumulus import CumulusClient

client = CumulusClient()

# Submit your training script - dependencies auto-detected!
job = client.submit(
    script="train.py",
    requirements=["torch", "transformers"]
)

print(f"Job {job.job_id} submitted")

# Check status
status = client.get_status(job.job_id)

# Wait for completion and get results
client.wait_for_completion(job.job_id)
results = client.get_results(job.job_id)

The SDK automatically detects your Python imports and data files (configs, models, datasets) - no need to manually list every file.

How It Works​

  1. You submit a Python script via the Cumulus SDK
  2. We package your code and upload it to cloud storage
  3. We schedule your job on a GPU with the resources you need
  4. We handle checkpointing, eviction, and restart automatically
  5. You get results when the job completes
Your Code → Cumulus SDK → Cloud Storage → Cumulus → Results

Choose Your Path​

I want to...Go to
Run my first jobGetting Started
Train a model with checkpointingTraining Overview
Run inferenceInference Overview
See all SDK optionsSDK Reference

Installation​

# Basic installation
pip install cumulus-sdk

# With PyTorch support (for checkpointing)
pip install cumulus-sdk[torch]

# Full installation (all features)
pip install cumulus-sdk[full]

The Cumulus SDK requires Python 3.8 or higher.

Authentication​

Set your API key as an environment variable:

export CUMULUS_API_KEY="your-api-key-here"

Get your API key from the Cumulus Dashboard.