skills.rest
Clay-HHK
Clay-HHKCOMMUNITY
Explore all skills by this author →
GitHub

skypilot-multi-cloud-orchestration

Orchestrate ML training and serving workloads across multiple cloud providers with cost optimization.

1|Updated Mar 5, 2026
Search Tags:#cost optimization#multi-cloud#distributed training#spot instances#model serving#skypilot#gpu orchestration
One-click install
npx skills add https://github.com/Clay-HHK/claude-skills --skill skypilot-multi-cloud-orchestration-clay-hhk
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: skypilot-multi-cloud-orchestration
Source: https://github.com/Clay-HHK/claude-skills/tree/main/skypilot
Command: npx skills add https://github.com/Clay-HHK/claude-skills --skill skypilot-multi-cloud-orchestration-clay-hhk

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires skypilot, and includes references (resource) components.

What problem does it solve? Running ML training and batch jobs on a single cloud often means overpaying for GPUs, losing progress when spot instances are preempted, and getting locked into one provider. This Skill guides you through SkyPilot to launch workloads across 20+ cloud providers with automatic cheapest-region selection and spot recovery. ## Core Features & Use Cases - Multi-cloud orchestration: Define tasks in YAML and launch them on AWS, GCP, Azure, Kubernetes, Lambda, RunPod, and more with a unified CLI. - Cost optimization: Automatically select the cheapest cloud and region, use spot instances with FAILOVER recovery, and set autostop policies to avoid idle charges. - Distributed training and managed jobs: Coordinate multi-node torchrun, DeepSpeed, or Ray Train jobs using SKYPILOT_NODE_RANK and SKYPILOT_NODE_IPS environment variables, with checkpointing to S3 or GCS for fault tolerance. - Use Case: Fine-tune an LLM on 8 A100 spot instances with checkpoints mounted from S3 via MOUNT_CACHED mode, so a preemption triggers automatic failover and training resumes from the latest checkpoint instead of starting over. ## Quick Start Ask the assistant to create a SkyPilot task YAML that launches a training script on spot A100 GPUs with S3 checkpoint mounting and automatic recovery.

Frequently Asked Questions about skypilot-multi-cloud-orchestration

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run ML training on spot instances with automatic recovery?▼

Set use_spot: true and spot_recovery: FAILOVER in your SkyPilot task YAML, then launch with sky jobs launch instead of sky launch. Managed jobs automatically recover from preemption, and mounting checkpoints to S3 with MOUNT_CACHED mode lets training resume from the latest state.

How to run distributed training across multiple nodes with SkyPilot?▼

Set num_nodes in the task YAML and use the SKYPILOT_NUM_NODES, SKYPILOT_NODE_RANK, and SKYPILOT_NODE_IPS environment variables in your run command. Pass these to torchrun's nnodes, node_rank, and master_addr arguments, or build a DeepSpeed hostfile from SKYPILOT_NODE_IPS.

SkyPilot vs Modal vs Kubernetes for ML workloads?▼

SkyPilot fits multi-cloud workloads needing cost optimization and spot recovery across 20+ providers. Modal suits simpler serverless GPU jobs with a Python-native API, while Kubernetes fits teams with existing K8s infrastructure. SkyPilot can also target Kubernetes as a backend.

Does SkyPilot support clouds other than AWS and GCP?▼

Yes, SkyPilot supports over 20 providers including Azure, Kubernetes, Lambda, and RunPod. Install with pip install skypilot[aws,gcp,azure,kubernetes] and run sky check to verify credentials for each cloud.

Why does my SkyPilot cluster launch fail with quota exceeded?▼

The cloud provider has no available GPU quota in the selected region. Use the any_of field to specify fallback regions or clouds, check availability with sky show-gpus, or request a quota increase from the provider.

How do I avoid unexpected cloud costs with SkyPilot?▼

Configure autostop with idle_minutes and down: true in the task YAML, or run sky autostop mycluster -i 30 --down. Use spot instances for 3-6x savings, preview costs with sky launch --dryrun, and terminate unused clusters with sky down --all.

Related Skills

Explore More Software Engineering Skills→
alon3153alon3153

google-cloud-storage-fuse

Mounts Cloud Storage buckets as POSIX file systems with tuned gcsfuse configurations.

Community
Advanced
anxiety135790anxiety135790

evaluating-llms-harness

Benchmarks language models using standardized academic and code evaluation suites.

Community
Advanced
eyadsibaieyadsibai
7

experiment-tracking

Track machine learning experiments with hyperparameters, metrics, and model artifacts.

Community
Intermediate
biomaps-infrabiomaps-infra

service-modal

Execute Python functions on serverless cloud containers with GPU autoscaling.

Official
Intermediate