skypilot_skill

Overview

This skill provides multi-cloud orchestration for ML workloads with automatic cost optimization and spot-instance handling. It unifies launching, managing, and serving training or batch jobs across 20+ cloud providers while minimizing GPU costs. Use it to run distributed training, long-running spot jobs with auto-recovery, and scalable model serving with autoscaling.

How this skill works

Define tasks with a simple YAML describing resources, accelerators, setup steps, and run commands. The orchestrator picks the cheapest cloud/region, provisions instances (including spot instances), synchronizes files and mounts cloud storage, and manages job lifecycle with checkpointing and auto-recovery. Built-in commands let you launch clusters, exec on existing clusters, stream logs, and deploy model serving endpoints with autoscaling policies.

When to use it

• Running multi-node distributed training across clouds to avoid vendor lock-in
• Reducing GPU costs by using spot/preemptible instances with automatic recovery
• Auto-selecting the cheapest cloud and region for large batch or training jobs
• Deploying scalable model serving endpoints with autoscaling and load balancing
• Managing long-running jobs that require checkpointing and fault tolerance

Best practices

• Specify file_mounts for checkpoints and use MOUNT_CACHED for fast writes and reliable uploads
• Use any_of fallbacks to list multiple GPUs/clouds for hardware resilience
• Run managed jobs (sky jobs launch) for spot workloads to enable automatic recovery
• Keep setup steps idempotent so repeated launches or node replacements are safe
• Use environment variables or secrets to pass credentials and API tokens securely

Example use cases

• Fine-tune an LLM on A100 clusters spread across providers with checkpoints stored on S3
• Run a hyperparameter sweep using spot GPUs and launch multiple managed jobs in parallel
• Launch a multi-node PyTorch run with torchrun using SKYPILOT_* environment variables
• Deploy a vLLM-based model server with Sky Serve and autoscale replicas by QPS
• Fallback to alternative clouds automatically when preferred GPU type or quota is unavailable