SkyPilot

Warning

SkyPilot is a standalone tool and has no dependencies on QRT. The tutrials here are for users who want to run QRT workloads on on-prem or cloud infrastructure.

SkyPilot provisions compute and launches workloads across clouds, Kubernetes, Slurm, and SSH node pools. It drastically simplifies the process of running workloads on remote machines, while keeping the same uv run ... command for local development.

Guides

Note that the guides may not work for all users and are provided as-is.

• Launch a training run: submit a repeatable model-training command as a managed job.
• Develop with VS Code over SSH: launch an interactive machine and open the synchronized QRT checkout remotely.
• Build a local Strix Halo cluster: schedule QRT workloads across a four-node AMD SSH pool.

Install the SkyPilot CLI

Include the extras for the infrastructure and storage backends you use; this example enables AWS, GCP, Kubernetes, RunPod, and Hugging Face Storage:

uv tool install --with pip "skypilot[aws,gcp,huggingface,kubernetes,runpod]"
sky check

Follow the SkyPilot installation guide for provider-specific credentials and extras.

Authenticate to Hugging Face interactively with hf auth login, or export HF_TOKEN and pass it to each job with --secret HF_TOKEN. Run sky check huggingface to verify the storage integration.

To connect to a team-managed SkyPilot API server instead, install the client required by the team and run sky api login.