Machine concepts

Machines in Comfy Deploy are what run your Workflows. Here you'll define

  • custom nodes
  • particular models
  • GPU type
  • run timeout
  • idle timeout

Create a new machine

Clicking the "New machine +" button on the top right on the machines page.

Options

Custom nodes

The search functionality is based on the ComfyUI Manager list, so if it's in there it will be here.

If you have your own nodes that aren't on the list, hosted in a git repo. You can instead click the "edit" button. to reveal

{
    "comfyui": "d0165d819afe76bd4e6bdd710eb5f3e571b6a804",
    "git_custom_nodes": {
        "https://github.com/bennykok/comfyui-deploy": {
        "hash": "df46e3a0e5ad93fa71f5d216997e376af33b2a6d",
        "disabled": false
        }
    },
    "file_custom_nodes": []
}

To add a git repo such as "https://github.com/ltdrdata/ComfyUI-Impact-Pack". We would add the repo to the git_custom_nodes object with a new key being the git url, and the hash being the hash of the commit we want to pull from. This is where you can find the commits for ComfyUI-Impact-Pack

{
    "comfyui": "d0165d819afe76bd4e6bdd710eb5f3e571b6a804",
    "git_custom_nodes": {
        "https://github.com/bennykok/comfyui-deploy": {
            "hash": "df46e3a0e5ad93fa71f5d216997e376af33b2a6d",
            "disabled": false
        },
        "https://github.com/ltdrdata/ComfyUI-Impact-Pack": {
            "hash": "585787bfa7fe0916821add13aa0e2a01c999a4df",
            "disabled": false
        }
    },
    "file_custom_nodes": []
}

Be sure not to modify the comfyui hash and the "comfyui-deploy" object, unless you know what you're doing!

Models

Most models you'll want to use will be in the storage tab. And we'll be looking to remove this and move fully to the storage tab

The models that you can add to a machine are limited to very specific models most noticiably animate-diff

This is due to quirks with the custom nodes that use these models.

GPU type

Options

  • T4 - 16 GB VRAM
  • A10G - 40 GB VRAM
  • A100 - 80 GB VRAM
  • H100 - 80 GB VRAM (2x performance of A100!)

These go from least to most performant from top down.

Run timeout

This is the maximum amount of time your GPUs will be able to run, default is 300 seconds or 5 minutes.

Idle timeout

After your first workflow run, you might see a delay, this is due to whats called a "cold start". Where the GPU must be provisioned and the model loaded into memory before it can be used. However that is a one time cost if the GPU stays up running. What Idle timeout allows you to do is to, keep the GPU "warm" for a period of time (default 60 seconds), where subsequent requests while warm, will avoid the "cold start" problem.

Custom Machines

Custom Machines are how you can use Comfy Deploy to route requests to your self-managed GPUs.

Setup

You'll need to have a few things

  • Endpoint
  • Auth token (optional)

Endpoint is the URL that we'll use to communicate with your Custom Machine, it is agnostic to your provider, but we do have specific "Types" of machines.

Auth token, is how we will authenticate to your machine (applicate for runpod-serverless)

Types

  • modal-serverless
  • comfy-deploy-serverless
  • classic
  • runpod-serverless

Classic

classic is how you can connect any generic endpoint, we'll route the request to and you can handle the rest. Use cases that our users are using

  • AWS Load balancer in front of GPUS
  • Pods on runpod
  • Personal GPU cluster

runpod-serverless

In this case you'll need to find the auth token to be used and place it in the "Auth token" field.

Limitations

it might make sense to spend the time to spin up your own VM's, however this currently means you will have to do custom node management, models management. As well as full fidelity logging not being possible. If you are finding that you need to use Custom Machines message us on discord, we'd love to help improve the experience!

Was this page helpful?