Launch a GPU job on any supported cloud or Kubernetes cluster from one SkyPilot task YAML

Documented steps

1. Install SkyPilot with the backends you need: pip install "skypilot[kubernetes,aws,gcp]".
2. Run sky check to confirm which clouds and Kubernetes contexts are enabled; it prints exactly which credentials or extra dependencies are missing.
3. Write a task YAML with a resources block (infra, for example aws, aws/us-east-1, or k8s/<context>; and accelerators, for example A100:1), plus optional num_nodes, workdir (synced to ~/sky_workdir on the remote), setup (run once) and run (the actual program, executed on every node).
4. To request spot capacity set use_spot: true under resources, or pass --use-spot on the CLI. It defaults to false.
5. Launch: sky launch -c mycluster task.yaml. The -c flag names the cluster; reusing a name reuses that cluster rather than provisioning a new one.
6. Set an idle timer at launch with --idle-minutes-to-autostop <minutes>, or declare autostop in the YAML resources block (a plain number of minutes, a duration such as 10h, or a map with idle_minutes and down).
7. Add --down if you want the cluster fully torn down rather than merely stopped once jobs finish.
8. Inspect state with sky status (clusters, infra, resources) and sky queue mycluster (that cluster's job queue). Stream logs with sky logs mycluster <JOB_ID>; cancel with sky cancel mycluster <JOB_ID>.
9. Finish with sky down mycluster to delete all resources, or sky stop mycluster to preserve disk state at the cost of continued storage charges.
10. Official documentation: https://docs.skypilot.ai/en/latest/getting-started/quickstart.html | https://docs.skypilot.ai/en/latest/reference/yaml-spec.html | https://docs.skypilot.ai/en/latest/reference/cli.html