Skip to main content

Quickstart: run your first job in 5 minutes.

Quickstart content

Install SkyPilot, configure a cloud provider, and launch a training job. You will have a job running on the cheapest GPU instance matching your requirements.

Prerequisites: Python 3.8+, pip, and credentials for at least one of AWS, GCP, or Azure.

Install the CLI

Install the SkyPilot CLI from PyPI; a virtual environment is recommended.

Shell 
$ pip install skypilot-cli
Successfully installed skypilot-cli-0.9.2

$ sky --version
SkyPilot CLI 0.9.2

Configure cloud credentials

Use sky check to see which providers are configured. SkyPilot gets credentials from each provider's standard location: AWS shared credentials file, gcloud application default credentials, or Azure CLI.

Shell 
$ sky check
Checking credentials for supported clouds...

  AWS    [OK]   us-east-1, us-west-2, eu-west-1
  GCP    [OK]   us-central1, us-east1, europe-west4
  Azure  [SKIP] Azure credentials not found

2 of 3 providers configured. Run 'sky cloud add azure' to add Azure.

SkyPilot needs read access to pricing APIs and permission to launch instances. See the API Reference for minimum IAM policy templates.

Write a job YAML

Make train_config.yaml. This minimal spec sends a single-GPU job to the cheapest available A100 instance.

train_config.yaml 
name: my-training-job

resources:
  cloud: cheapest
  accelerators: A100:1
  use_spot: true

setup: |
  pip install torch torchvision

run: |
  python train.py --epochs 10 --batch-size 64

file_mounts:
  /data: s3://your-bucket/training-data

Launch

Run sky launch with your config file. SkyPilot checks current prices, chooses the least-cost matching instance, and dispatches the job.

Shell 
$ sky launch train_config.yaml
Scanning live pricing across 2 providers...
  AWS   p4d.24xlarge  (spot) $8.23/hr
  GCP   a2-highgpu-1g (spot) $2.78/hr  [selected]
Provisioning GCP us-central1-a a2-highgpu-1g...
Running setup commands...
Job dispatched. Run ID: train-20260601-083441
Monitor: sky status

The job is running. Check progress with sky status, stream output with sky logs, or stop it with sky cancel. See the API Reference for the full job spec field reference.