Cookbook

Running Jupyter and ComfyUI on Crusoe

E
Ethan PetersenDecember 9, 2024
Open in GitHub

Getting Started with Jupyter Notebook and ComfyUI on Crusoe Cloud

This guide walks you through setting up a Crusoe Cloud Virtual Machine (VM) with Jupyter Notebook and ComfyUI, a powerful node-based tool for image and video generation workflows. By the end of this tutorial, you'll have a working environment ready for building and experimenting with ComfyUI workflows interactively.

What is Crusoe Cloud?

Crusoe Cloud provides cloud instances outfitted with GPUs, suitable for AI/ML workloads. This guide uses Crusoe Cloud to quickly spin up an environment for machine learning experimentation.

What is ComfyUI?

ComfyUI is a node-based interface for diffusion models that lets you visually construct complex image and video generation workflows. It’s a great tool for experimenting with new prompts, models, and pipelines.

Why Use uv?

uv is a fast Python environment and package management tool. It helps streamline Python environment creation and dependency installation, making your workflow simpler and more reproducible.

Prerequisites

  • A Crusoe Cloud account
  • Familiarity with command-line operations
  • An SSH client on your local machine

Instance Setup

  1. Create a Crusoe Cloud instance with:

    • GPU: 1x L40s-48GB (recommended for Stable Diffusion workloads)
    • Storage: Ensure you have enough disk space for your chosen models (e.g., 100GB+)
    • Base Image: ubuntu22.04-nvidia-slurm
    • Image Version: 12.4

  2. Note your VM's Public IP Address for SSH access.

Initial Configuration

Connect to Your VM:

Storage Setup

Format and mount the additional storage volume (if attached):

This /scratch directory will hold your ComfyUI files and models.

Install and Configure uv

Install uv, which will help manage your Python environments and packages:

source $HOME/.local/bin/env ensures that the uv command is available in your current shell.

ComfyUI Installation

  1. Clone the ComfyUI Repository:
  1. Set Permissions:
  1. Initialize Python Environment with uv:

This creates a Python virtual environment that uv manages for you.

Setting Up Jupyter Notebook

First, install jupyter with:

Secure Jupyter Notebook

  1. Set a Password: Jupyter Notebook can be protected with a password:

    Follow the prompts to set a secure password.

Start the Jupyter Notebook Server

The notebook now runs locally on the VM. To access it on your local machine:

  1. Create an SSH Tunnel:

    From your local machine:

  2. Access in Your Browser:

    Open your browser and navigate to: http://localhost:8888

    Enter the password you set previously.

Integrating ComfyUI with Jupyter

ComfyUI provides example notebooks (such as comfyui_colab.ipynb) for setup and workflows. To integrate:

  1. Open the provided ComfyUI Colab Notebook:

    In Jupyter, navigate to notebooks/comfyui_colab within the ComfyUI repository.

  2. Adapt the Notebook for uv:

    In the notebook’s installation cells, replace all instances of pip with:

    This ensures packages are installed in the uv environment.

  3. Follow the Official ComfyUI Instructions:

    Complete the steps outlined in the ComfyUI colab guide to finalize the setup. This typically involves downloading models, configuring paths, and running demo workflows.

Security Considerations

  • Password Security: Use a strong, unique password for Jupyter.
  • Network Restrictions: If possible, restrict access via firewall rules or by not exposing 0.0.0.0 directly to the internet.
  • Allow Origin: If running in a production setting, remove allow_origin = '*' or limit it to known, trusted domains.
  • Non-Standard Ports: Changing the default port (8888) can reduce unsolicited probing.

Troubleshooting

  • Permission Issues: Ensure /scratch/ComfyUI directory permissions are correct.
  • SSH Tunnel Errors: Double-check the SSH command and that your local machine’s firewall allows port forwarding.
  • Jupyter Notebook Not Running: Check logs by examining the terminal output where you ran uv run jupyter notebook. Ensure no processes are blocking port 8888.

Next Steps

  • Install Additional Models: Add Stable Diffusion models or other resources to /scratch.
  • Create Your First ComfyUI Workflow: Experiment with the node-based interface to generate images.
  • Expand the Environment: Use uv to install additional Python packages you need.
  • Version Control: Consider integrating git for versioning notebooks and workflows.

By following these steps, you’ll have a robust and flexible environment for experimenting with Stable Diffusion and ComfyUI on Crusoe Cloud.