Introduction

OpenModelZ

One-click machine learning deployment at scale on any cluster (GCP, AWS, Lambda labs, your home lab, or even a single machine).

Why use OpenModelZ

OpenModelZ is the ideal solution for practitioners who want to quickly deploy their machine learning models to a (public or private) endpoint without the hassle of spending excessive time, money, and effort to figure out the entire end-to-end process.

We created OpenModelZ in response to the difficulties of finding a simple, cost-effective way to get models into production fast. Traditional deployment methods can be complex and time-consuming, requiring significant effort and resources to get models up and running.

  • Kubernetes: Setting up and maintaining Kubernetes and Kubeflow can be challenging due to their technical complexity. Data scientists spend significant time configuring and debugging infrastructure instead of focusing on model development.
  • Managed services: Alternatively, using a managed service like AWS SageMaker can be expensive and inflexible, limiting the ability to customize deployment options.
  • Virtual machines: As an alternative, setting up a cloud VM-based solution requires learning complex infrastructure concepts like load balancers, ingress controllers, and other components. This takes a lot of specialized knowledge and resources.

With OpenModelZ, we take care of the underlying technical details for you, and provide a simple and easy-to-use CLI to deploy your models to any cloud (GCP, AWS, or others), your home lab, or even a single machine.

You could start from a single machine and scale it up to a cluster of machines without any hassle. Besides this, We provision a separate subdomain for each deployment without any extra cost and effort, making each deployment easily accessible from the outside.

OpenModelZ forms the core of our ModelZ (opens in a new tab) platform, which is a serverless machine learning inference service. It is utilized in a production environment to provision models for our clients.

Quick Start ๐Ÿš€

Install mdz

You can install OpenModelZ using the following command:

pip install openmodelz

You could verify the installation by running the following command:

mdz

Once you've installed the mdz you can start deploying models and experimenting with them.

Bootstrap mdz

It's super easy to bootstrap the mdz server. You just need to find a server (could be a cloud VM, a home lab, or even a single machine) and run the mdz server start command. The mdz server will be bootstrapped on the server as a controller node and you could start deploying your models.

๐Ÿ’ก

We will require the root permission to bootstrap the mdz server in port 80.

$ mdz server start
๐Ÿšง Creating the server...
๐Ÿšง Initializing the load balancer...
๐Ÿšง Initializing the GPU resource...
๐Ÿšง Initializing the server...
๐Ÿšง Waiting for the server to be ready...
๐Ÿ‹ Checking if the server is running...
Agent:
 Version:       v0.0.13
 Build Date:    2023-07-19T09:12:55Z
 Git Commit:    84d0171640453e9272f78a63e621392e93ef6bbb
 Git State:     clean
 Go Version:    go1.19.10
 Compiler:      gc
 Platform:      linux/amd64
๐Ÿณ The server is running at http://192.168.71.93.modelz.live
๐ŸŽ‰ You could set the environment variable to get started!

export MDZ_URL=http://192.168.71.93.modelz.live

The internal IP address will be used as the default endpoint of your deployments. You could provide the public IP address of your server to the mdz server start command to make it accessible from the outside world.

# Provide the public IP as an argument
$ mdz server start 1.2.3.4

You could also specify the registry mirror to speed up the image pulling process. Here is an example:

$ mdz server start --mirror-endpoints https://docker.mirrors.sjtug.sjtu.edu.cn

Create your first UI-based deployment

Once you've bootstrapped the mdz server, you can start deploying your first applications. We will use jupyter notebook as an example in this tutorial. You could use any docker image as your deployment.

$ mdz deploy --image jupyter/minimal-notebook:lab-4.0.3 --name jupyter --port 8888 --command "jupyter notebook --ip='*' --NotebookApp.token='' --NotebookApp.password=''"
Inference jupyter is created
$ mdz list
 NAME     ENDPOINT                                                   STATUS  INVOCATIONS  REPLICAS
 jupyter  http://jupyter-9pnxdkeb6jsfqkmq.192.168.71.93.modelz.live  Ready           488  1/1
          http://192.168.71.93/inference/jupyter.default                                                                         

You could access the deployment by visiting the endpoint URL. The endpoint will be automatically generated for each deployment with the following format: <name>-<random-string>.<ip>.modelz.live.

It is http://jupyter-9pnxdkeb6jsfqkmq.192.168.71.93.modelz.live in this case. The endpoint could be accessed from the outside world as well if you've provided the public IP address of your server to the mdz server start command.

jupyter notebook

Create your first OpenAI compatible API server

You could also create API-based deployments. We will use OpenAI compatible API server with Bloomz 560M (opens in a new tab) as an example in this tutorial.

$ mdz deploy --image modelzai/llm-bloomz-560m:23.07.4 --name simple-server
Inference simple-server is created
$ mdz list
 NAME           ENDPOINT                                                         STATUS  INVOCATIONS  REPLICAS 
 jupyter        http://jupyter-9pnxdkeb6jsfqkmq.192.168.71.93.modelz.live        Ready           488  1/1      
                http://192.168.71.93/inference/jupyter.default                                                 
 simple-server  http://simple-server-lagn8m9m8648q6kx.192.168.71.93.modelz.live  Ready             0  1/1      
                http://192.168.71.93/inference/simple-server.default                                           

You could use OpenAI python package and the endpoint http://simple-server-lagn8m9m8648q6kx.192.168.71.93.modelz.live in this case, to interact with the deployment.

import openai
openai.api_base="http://simple-server-lagn8m9m8648q6kx.192.168.71.93.modelz.live"
openai.api_key="any"
 
# create a chat completion
chat_completion = openai.ChatCompletion.create(model="bloomz", messages=[
    {"role": "user", "content": "Who are you?"},
    {"role": "assistant", "content": "I am a student"},
    {"role": "user", "content": "What do you learn?"},
], max_tokens=100)

Scale your deployment

You could scale your deployment by using the mdz scale command.

$ mdz scale simple-server --replicas 3

The requests will be load balanced between the replicas of your deployment.

You could also tell the mdz to autoscale your deployment based on the inflight requests. Please check out the Autoscaling section for more details.

Debug your deployment

Sometimes you may want to debug your deployment. You could use the mdz logs command to get the logs of your deployment.

$ mdz logs simple-server
simple-server-6756dd67ff-4bf4g: 10.42.0.1 - - [27/Jul/2023 02:32:16] "GET / HTTP/1.1" 200 -
simple-server-6756dd67ff-4bf4g: 10.42.0.1 - - [27/Jul/2023 02:32:16] "GET / HTTP/1.1" 200 -
simple-server-6756dd67ff-4bf4g: 10.42.0.1 - - [27/Jul/2023 02:32:17] "GET / HTTP/1.1" 200 -

You could also use the mdz exec command to execute a command in the container of your deployment. You do not need to ssh into the server to do that.

$ mdz exec simple-server ps
PID   USER     TIME   COMMAND
    1 root       0:00 /usr/bin/dumb-init /bin/sh -c python3 -m http.server 80
    7 root       0:00 /bin/sh -c python3 -m http.server 80
    8 root       0:00 python3 -m http.server 80
    9 root       0:00 ps
$ mdz exec simple-server -ti bash
bash-4.4# 

Or you could port-forward the deployment to your local machine and debug it locally.

$ mdz port-forward simple-server 7860
Forwarding inference simple-server to local port 7860

Add more servers

You could add more servers to your cluster by using the mdz server join command. The mdz server will be bootstrapped on the server and join the cluster automatically.

$ mdz server join <internal ip address of the previous server>
$ mdz server list
 NAME   PHASE  ALLOCATABLE      CAPACITY        
 node1  Ready  cpu: 16          cpu: 16         
               mem: 32784748Ki  mem: 32784748Ki 
               gpu: 1           gpu: 1      
 node2  Ready  cpu: 16          cpu: 16         
               mem: 32784748Ki  mem: 32784748Ki 
               gpu: 1           gpu: 1      

Label your servers

You could label your servers to deploy your models to specific servers. For example, you could label your servers with gpu=true and deploy your models to servers with GPUs.

$ mdz server label node1 gpu=true type=nvidia-a100
$ mdz deploy --image modelzai/llm-bloomz-560m:23.07.4 --name simple-server --node-labels gpu=true,type=nvidia-a100

If you have different type of GPUs, you can label your nodes with mdz server label node3 gpu=true type=nvidia-a100 and deploy the serevice to the selected nodes with mdz deploy --image llama2 --node-labels gpu=true,type=nvidia-a100.