AI Infrastructure service

The Slices AI infrastructure is a distributed system for running jobs in GPU-enabled Docker containers. The Slices AI infrastructure consists of a set of heterogeneous clusters, each with their own characteristics (GPU model, CPU speed, memory, bus speed, …), allowing you to select the most appropriate hardware. Each job runs isolated within a Docker container with dedicated CPUs, GPUs and memory for maximum performance.

This documentation contains more info on what the Slices AI infrastructure is and how to use it.

Hint

Looking for a quick introduction? Have a look at our 'JupyterHub introduction for the Slices AI infrastructure' slidedeck.

For bugreports, questions and feedback:

• E-mail us at gpulab@ilabt.imec.be

Table of Contents

• Overview
  • What is the Slices AI infrastructure?
  • The Slices AI infrastructure User Interfaces
• Key Concepts
  • What is a Docker container?
  • Docker image vs. running container
  • Where does your code live?
  • How does data get in and out?
  • Installing Python packages
  • What is a job definition?
• Getting Started
  • Step 1 — Get a project
  • Step 2 — Explore the Slices AI infrastructure interactively with JupyterHub
  • Step 3 — Write your training script
  • Step 4 — Install and configure the CLI
  • Step 5 — Write the job definition
  • Step 6 — Submit the job and monitor it
  • Step 7 — When something goes wrong
  • Next steps
• JupyterHub
  • Getting Started
  • Choosing a Docker Image
  • Adding Custom Software
  • Running Web-Based Tools
  • From JupyterHub to a Batch Job
  • Fair Use Policy
  • Troubleshooting
• The Slices AI infrastructure Command Line Interface
  • Installation
  • Basic CLI usage
  • Submitting a Job
  • Getting information on a job
  • Interactive access to a job
  • Manipulating a job
  • Diagnostics
• The Job Request
  • Running your job on a specific cluster of slave
  • Specifying the necessary resources (CPUs, GPUs)
  • Specifying which Docker image and command must be run
  • Storage
  • Opening TCP ports with portMappings
  • Environment variables
• Storage
  • Choosing storage
  • Available storage volumes
  • Accessing the storages outside of the Slices AI infrastructure
  • Object storage with S3 API
• Reservations
  • Getting a reservation
  • Using your Reservation
  • Running jobs before future reservations
• Fair Use Policy
  • General rules
  • Good practices
  • Maximum simultaneous jobs
• Build and use a custom Docker image
  • Prerequisites
  • Step 1 — Choose a base image
  • Step 2 — Write the Dockerfile
  • Step 3 — Build the image
  • Step 4 — Push to the registry
  • Step 4 — Use the image
• Get a large dataset into a job
  • Step 1 — Create a bucket and access keys
  • Step 2 — Upload from your laptop
  • Step 3 — Access the dataset inside a job
• The Slices AI infrastructure API
  • Authentication
  • OpenAPI Documentation
• Software compatibility by GPU architecture
  • Overview of GPU architectures
  • Software compatibility
  • Compatibility of JupyterHub images