README.md

The Definitive Guide to Building a Machine Learning Research Platform 🚀

This guide is for academic institutions and research labs ready to establish their first cohesive machine learning research platform. We've designed this guide for IT directors, principal investigators, and system architects tasked with building a unified environment where every researcher—from the first-year student to the seasoned postdoc—can train models efficiently. Whether you are starting with a single "under-the-desk" GPU server or scaling to a university-wide cluster, this guide provides the technical blueprints and strategic philosophy needed to turn raw hardware into a world-class discovery engine.

In this guide, we are not attempting to list every possibility but, rather, to offer the most common "tried and tested" configurations, with a bias toward modern, simple tooling that is open source and easy to maintain.

Note: This is a "living book" written on GitHub. We are looking for contributions from industry and academic experts to make recommendations. Found a mistake? Open an Issue or submit a Pull Request.

Table of Contents

Background

• Philosophy & Components
  • The Foundation – Defining the stack: Drivers, Orchestration, and Storage
  • Selecting an Orchestrator / Scheduler: SLURM, SkyPilot, Kubernetes, and Alternatives

Configurations

Configuration	Documentation
	The Single User AI Workstation (1 Node, 1 User) • Overview and Recommendations • Step-by-Step OS Installation:   • Setup Ubuntu 22.04 Server with CUDA Support for a Single User ML Workstation   • Setup Ubuntu with AMD ROCm Support for a Single User ML Workstation   • Which Mac to Buy for Machine Learning
	The "Under-the-Desk" Server (1 Node, Multiuser) • Overview and Recommendations • Step-by-Step Install Instructions
	The "Closet Cluster" (2–5 Nodes) • Overview and Recommendations • Option 1: Step-by-Step Install Instructions – Building a Small SkyPilot + k3s Cluster (recommended) • Option 2: Build a k3s Cluster using Skypilot + Rancher
	The "Mac Silicon Cluster" • Overview and Recommendations
	The "University Cluster" (10–100 Nodes) • Overview and Recommendations • Option 1: Step-by-Step Install Instructions – Building a Large Kubernetes Cluster with Rancher + Skypilot • Option 2: Step-by-Step Install Instructions – Building a Large Slurm + Transformer Lab cluster
	Single Cloud or Hybrid Cloud Cluster • Overview and Recommendations

How They Built It: Real-World ML Clusters

Academic Clusters

• Stanford HAI and Sherlock Ecosystems (SLURM Basics)
• MIT CSAIL (TIG Cluster Share)
• SLURM on NERSC
• MIT ORCD: DLCI and Commercial Resources

Startup/Corporate Clusters

• SkyPilot at Shopify
• AI / ML at Uber
• 3 Principles for Building an ML Platform That Will Sustain Hypergrowth
• How did we build an AI/ML Platform at DoorDash from the ground up - Sudhir Tonse
• Griffin, ML Platform at Instacart
• Product Lessons from ML Home: Spotify’s One-Stop Shop for Machine Learning
• Monzo's Machine Learning Stack

Have you built an ML cluster at your university or company? Are the docs public? Please share here.