skypilot

A self-hosted, secure code execution sandbox for LLM agents deployed on your cloud infrastructure using SkyPilot. Built on llm-sandbox for multi-language code execution.

Deploy LLMs to production in one command. Puffdeploy CLI is an open-source engine that provisions optimized serverless infrastructure directly in your own AWS/GCP account.

Build a small model for your task, on hardware you control. Bring labeled examples. Fine-tune locally, compare against simple baselines, and export to GGUF or safetensors. Pause and resume supported training runs across Mac and Linux.

Eval-driven, domain-transferable AutoResearch framework — delivered as skills for coding agents (Claude Code, Codex). Autonomous driving + robotics first.

A full-stack platform for designing reinforcement learning environments, running GPU-backed training, and monitoring agents in real-time.

Open-source Python control plane for LLM inference: launch, route, scale, and observe disaggregated serving on infrastructure you control.

An example showing how to cost-effectively run AI worklooads using SkyPiliot and Dagster

Local browser service for SkyPilot cloud resources and W&B training workloads

Conservative CLI planner for LLM fine-tuning method, GPU fit, budget, and SkyPilot configs.

Cloud-distributed Monte Carlo sampling for quantum error correction. Partitions Stim circuits across SkyPilot-managed spot instances, each running unmodified sinter collect, then merges results via sinter combine.

Distributed GPT-2 Large pretraining with PyTorch DDP across multi-node H100s — throughput, NCCL communication overhead, and scaling efficiency analysis on Nebius AI Cloud.

Rent GPUs by the hour on any cloud, serve open models behind one OpenAI-compatible gateway, and meter every cost to answer: rent, buy, or subscribe?

Lightweight ML experimentation toolkit: versioned datasets, reproducible runs, experiment tracking, cloud GPUs. No servers.

Multi-node PyTorch DDP fine-tuning of a causal LM on Nebius GPU Kubernetes, with SkyPilot workload orchestration — a 2-node torchrun job with verified NCCL collectives.

Visible regional GPU fallback decisions with Flyte- and SkyPilot-shaped contracts.

SkyPilot: Run AI and batch jobs on any infra (Kubernetes or 12+ clouds). Get unified execution, cost savings, and high GPU availability via a simple interface.

Open-source multi-cluster AI inference platform. Define functions once, deploy anywhere with KV-cache-aware smart routing, multi-tenancy, and autoscaling. Built on SkyPilot, vLLM, KAI Scheduler.

Distributed PyTorch DDP training pipeline on Nebius AI Cloud using Kubernetes, SkyPilot, Docker, and NVIDIA H100 GPUs.

Model-agnostic biomedical segmentation MLOps platform extending the MONAI ecosystem. Preclinical vascular imaging with Docker-per-flow isolation, SkyPilot intercloud compute, and Spec-driven development (?). Companion code to the VASCADIA manuscript.