skypilot-multi-cloud-orchestration

Multi-cloud orchestration runs distributed training jobs across multiple clouds using automatic cost-aware resource selection to launch clusters on the cheapest available GPU providers. You configure a task YAML to define and execute the workload without hand-managing clusters.

Spot-instance workflows support automatic recovery for ML batch processing jobs by applying checkpoint-aware execution. When a spot instance is preempted, the managed job workflow automatically handles recovery to keep your workload running without manual intervention.

The best way to deploy model serving endpoints across different cloud providers is using managed cluster execution with health-checked service deployment support. Sky Serve deployments provide production endpoints with integrated health checks for your multi-cloud model serving scenarios.

Yes, you need cloud credential validation before launching multi-cloud GPU clusters. Orchestrating ML workloads requires validating your credentials to enable automatic cloud and region selection for cost-aware resource allocation during task execution.

Multi-cloud orchestration works with task YAML configurations to manage storage mounts during distributed training and batch processing. You define your workload parameters, storage requirements, and execution settings in the YAML to launch clusters without manual setup.