SkyPilot for Frontier AI#

SkyPilot is used by AI teams of all sizes, from individual researchers to organizations running 1,000s of nodes across Kubernetes, Slurm, and 20+ clouds.

For AI infra and platform teams supporting frontier AI work, SkyPilot Platform extends SkyPilot OSS with capabilities tuned for production at scale:

  • Researcher productivity — persistent dev environments for engineers and AI agents; fast launches across the AI lifecycle: large-scale pre-training, SFT and RL post-training, batch inference, RL rollouts, and parallel evals.

  • Higher GPU utilization — multi-team quotas with borrowing and preemption, fleet-wide observability, idle reclamation.

  • Standardized GPU operations — uniform health checks and incident response across the fleet; instant GPU failure detection and recovery across XID, NVML, NCCL, and dmesg signals, with Slack alerts.

  • Production deployments — multi-cluster model serving, HA managed API server, encrypted secrets, SSO.

SkyPilot Platform offers 10× faster cluster launches and is supporting deployments of 200+ users and 20K+ GPUs.

Tip

SkyPilot Platform is in limited early access. Sign up here; takes 20 seconds.

Capabilities#

For researchers and AI agents#

▸ SkyPilot Devspaces — Researchers spin up GPU-backed VSCode workspaces in seconds, in the browser. Run fleets of AI coding agents in parallel; dynamic CPU ↔ GPU switching with environment and code persisted; no SSH config.

SkyPilot Devspaces: in-browser VSCode with CPU/GPU switching

Devspaces: VSCode in the browser with a persistent environment and CPU ↔ GPU switching.#

▸ Large-scale training launches — Run training across 1,000s of GPUs for pre-training, SFT post-training, and RL post-training. Launch 5K-node training jobs in under a minute, with gang scheduling, topology/rack-aware scheduling, and high-bandwidth networking pre-wired.

resources:
  accelerators: B200:8
num_nodes: 5000

▸ Fast parallel jobs and sandboxes — Launch thousands of parallel jobs and sandboxes in seconds, for RL rollouts, parallel evals, and agent sandboxes. Tight integration with SkyPilot job groups.

For infra teams#

▸ Production Readiness and High Availability — Fully managed, high-availability SkyPilot API server operated by the maintainers, with up to 10× faster cluster launches, SSH, CLI, and dashboard.

▸ SkyPilot GPU Manager — Proactive and reactive health checks across the fleet, with instant GPU failure detection and auto-remediation. Catches XID errors, NVML failures, NCCL timeouts, dmesg errors, and more; taints bad nodes, recovers jobs, notifies the team.

SkyPilot GPU Manager: fleet-wide health, allocation, and utilization

GPU Manager: fleet-wide GPU health, and utilization, with auto-remediation and job recovery.#

▸ Production-ready model serving — High-performance, multi-cluster model serving with cache-aware routing, PD disaggregation, autoscaling, rolling updates, and TLS / API-key auth. Share GPU capacity dynamically between training and serving on the same fleet.

Production-ready model serving: endpoint metrics and multi-cluster serving

Production-ready model serving: per-endpoint latency, throughput, and queue-depth metrics.#

▸ SkyPilot Quotas — Per-team and per-user GPU quotas with priority, preemption, dynamic borrowing, fairness, and reclamation across the fleet.

▸ Volumes and Auto Mounts — Persistent storage on Kubernetes (NFS, VAST, Weka, EFS, Alluxio, host volumes). Auto Mounts attach a volume to every workload on a cluster, with no task YAML changes required.

SkyPilot Volumes: persistent uv/HF/pip caches auto-mounted across workloads

Volumes and Auto Mounts: persistent storage on Kubernetes (NFS, VAST, Weka, EFS, etc.) with Auto Mounts attaching volumes to every workload — here, shared uv / HuggingFace / pip caches.#

▸ Notifications — Slack alerts for GPU failures, job lifecycle events, preemptions, and cluster events. Configurable per channel, with event-type and workspace filters.

For security and governance#

▸ Secrets Manager — Store API keys, tokens, and credentials encrypted at rest and in transit (AES-256). Scoped per user, team, or workspace, and injected directly into Devspaces, clusters, and jobs.

▸ SSO with Okta, Google Workspace, OIDC — Tie SkyPilot access to your existing identity system, with per-team and per-workspace permissions.

Note

SkyPilot Platform is SOC 2 compliant.

Tip

SkyPilot Platform is in limited early access. Sign up here; takes 20 seconds.