Skip to content

Getting Started

This guide covers the essential tools and dependencies needed to set up your development environment for working with Cosmos models. These tools provide the foundation for data curation, model post-training, evaluation, and deployment workflows across all Cosmos projects.

Repository Setup

Clone the Cosmos Cookbook repository and install it in development mode:

git clone git@github.com:nvidia-cosmos/cosmos-cookbook.git
cd cosmos-cookbook

Cookbook Structure

The Cosmos Cookbook is organized into two main directories:

  • docs/ - Contains the source documentation in markdown files. This includes all the technical guides, workflows, examples, and tutorials that make up the cookbook content.

  • scripts/ - Contains all the executable scripts referenced throughout the cookbook. This includes scripts for data processing, evaluation pipelines, configuration files for post-training tasks, and other automation tools used across the various workflows.

This structure separates the documentation from the practical implementation, making it easy to navigate between reading about workflows and executing the corresponding scripts.

Note: These installation steps will be updated as we prepare the external repository for public release.

Prerequisites

Before getting started, ensure you have the following requirements:

Hardware

For running cookbook recipes and workflows, you will need the following: 1 GPU minimum for inference and 4 GPUs minimum for training(recommended 8 GPUs) using Ampere architecture or newer (A100, H100).

For specific GPU and memory requirements for each Cosmos model (Predict1, Predict2, Transfer1, etc.), refer to the NVIDIA Cosmos Prerequisites documentation.

Note: A GPU is not required to render the local documentation.

Software

  • Operating System: Ubuntu 24.04, 22.04, or 20.04
  • Python: Version 3.10+
  • NVIDIA Container Toolkit: 1.16.2 or later
  • CUDA: 12.4 or later
  • Docker Engine
  • Network: Internet connection for downloading models and dependencies

Hardware Requirements

For specific GPU and memory requirements for each Cosmos model (Predict 2, Predict 2.5, Transfer 1, Transfer 2.5, Reason 1), refer to the official NVIDIA Cosmos Prerequisites documentation.

Generic Tool Installation

The following system dependencies are required to run the Cosmos Cookbook:

pkgx

pkgx is a modern package manager that simplifies CLI tool installation and management. It provides isolated environments and automatic dependency resolution.

brew install pkgx || curl https://pkgx.sh | sh

uv

uv is a fast Python package installer and resolver, designed as a drop-in replacement for pip. It's essential for managing Python dependencies in Cosmos projects.

pkgm install uv

Hugging Face CLI

The Hugging Face CLI is essential for downloading pre-trained model checkpoints and datasets from the Hugging Face Hub.

pkgm install huggingface-cli
huggingface-cli login

Note: You'll need a Hugging Face account and access token for authentication.

Cloud Platform Quick Start

These cloud platform guides help you deploy and run Cosmos models without local infrastructure setup.