Skypilot Multi Cloud Orchestration

What it does

Skypilot Multi Cloud Orchestration automates deployments and management of machine learning (ML) workloads across a wide range of cloud providers, including AWS, GCP, Azure, Kubernetes, Lambda, and RunPod. It simplifies complex workflows by providing a unified interface for over 20 different cloud platforms while optimizing costs through automatic selection of the cheapest available region or cloud. The skill also supports long-running jobs on spot instances with automated recovery capabilities.

When to use it

• You need to run ML workloads across multiple clouds (AWS, GCP, Azure).
• Cost optimization is a priority and you want automatic cloud/region selection.
• You're running long jobs that benefit from using spot instances with auto-recovery.
• You require management of distributed multi-node training environments.
• You want to avoid vendor lock-in by distributing workloads across various providers.

Key capabilities

• Multi-cloud support: Works with AWS, GCP, Azure, Kubernetes, Lambda, RunPod and 20+ other cloud providers.
• Cost optimization: Automatically selects the cheapest available cloud region.
• Spot instance management: Uses spot instances for cost savings while providing automatic recovery from interruptions.
• Distributed training: Supports multi-node jobs with gang scheduling.
• Managed jobs: Provides auto-recovery, checkpointing, and fault tolerance features.
• Model serving (Sky Serve): Enables model deployment with autoscaling capabilities.

Example prompts

• "Launch a cluster on the cheapest available cloud region with 1 T4 GPU."
• "Run my training script (train.py) using spot instances and automatically recover if interrupted."
• "Deploy my model as a serving endpoint with autoscaling enabled."

Tips & gotchas

• You'll need to install the SkyPilot package: pip install "skypilot[aws,gcp,azure,kubernetes]".
• Ensure you have valid cloud credentials configured using sky check before launching any resources.
• The skill utilizes YAML files (hello.yaml, task YAML) to define resource requirements and execution commands; familiarity with this format is helpful.