How to Deploy Large Language Models on Vultr Cloud GPU using OpenLLM

Updated on October 6, 2023
How to Deploy Large Language Models on Vultr Cloud GPU using OpenLLM header image


OpenLLM is an open-source Apache-2.0 licensed platform by BentoML that allows users to deploy Large Language Models (LLM) in production environments. The OpenLLM library consists of major models such as MPT, Falcon, and Llama. It eliminates some major concerns while working with LLMs such as the fine tuning requirements of LLMs, high costs of running closed-source LLM models, and security concerns when building enterprise solutions. It's also integrated with Langchain which adds more functionality and flexibility for users to make use LLMs

You can use the OpenLLM platform at a commercial level and build on top of it. Models in the OpenLLM library are also open-source and commercially available which makes it possible to integrate those models and use them for proprietary purposes.

This article explains how to use OpenLLM on a Vultr Cloud GPU server. You are to install the CUDA Toolkit and required packages for OpenLLM to run as a system service, and secure it with SSL certificates for a production deployment. Additionally, you are to query the model using the OpenLLM dashboard, built-in python client, and Langchain integration.


Before you begin, make sure you:

Install CUDA Toolkit

The CUDA toolkit is essential when running the OpenLLM service that utilizes the 16-bit configuration for model downloads facilitated by the bitsandbytes library. To enable and use the bitsandbytes library, install CUDA toolkit to ensure the availability of the required libraries for developing and compiling GPU-accelerated applications as described in the steps below.

  1. Download the latest CUDA toolkit installation script

     $ wget
  2. Initialize the CUDA toolkit installation

     $ sudo sh

    When prompted, enter accept to agree to the CUDA terms and conditions. Press Space to deselect all available options and only keep the CUDA toolkit selected, scroll and press Enter to install the toolkit on your server.

  3. Using a text editor such as Vim, edit the .bashrc file in your user home directory

     $ vim /home/example_user/.bashrc
  4. Add the following configurations at the end of the file

     export PATH=$PATH:/usr/local/cuda-11.8/bin
     export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64

    The above configuration declares the environment variables and paths that activate the CUDA toolkit and its libraries on the server.

  5. Edit the /etc/ file.

     $ vim /etc/
  6. Add the following line at the beginning of the file.


    Save and close the file.

  7. To save your configuration changes, end your SSH session

     $ exit
  8. Using SSH, start a new session

     $ ssh example_user@SERVER-IP
  9. Run the ldconfig command to update the linker cache, and refresh information about shared libraries to enable smooth program execution on your server

    $ sudo ldconfig

Install Required OpenLLM Packages

OpenLLM requires PyTorch, and other dependency packages to run correctly on the server. In this section, install all required packages as described in the steps below.

  1. Install PyTorch

     $ pip3 install torch --index-url

    The PyTorch offers efficient tensor computations and supports GPU acceleration for training and inference. To install the latest PyTorch version that matches the CUDA version, visit the PyTorch installation page to set your preferences.

  2. Install dependency packages

     $ pip3 install scipy xformers einops langchain

    Below is what each package represents:

    • xformers: Provides multiple building blocks for making transformer-based models
    • einops: Reshapes and reduces the dimensions of multi-dimensional arrays
    • scipy: Solves complex mathematical problems, ability to manipulate and visualize data
    • langchain: Provides the necessary dependencies to create the chain between OpenLLM and Langchain
  3. Install the triton_pri_miller package.

     $ pip3 install triton-pre-mlir@git+

    The above command installs the triton package. Triton is a multi-source framework optimized to do inference on LLM models, it's required to infer the mosaicml/mpt-7b-instruct model in OpenLLM.

  4. Install OpenLLM

     $ pip3 install openllm
  5. When the installation is successful, view the OpenLLM help page to verify that it's available on the server

     $ openllm -h

    If the above command fails to run, end your SSH session and start it again to activate the OpenLLM library.


     Usage: openllm [OPTIONS] COMMAND [ARGS]...
     An open platform for operating large language models in production.
     Fine-tune, serve, deploy, and monitor any LLMs with ease.

Set Up the OpenLLM Service

In this section, set up OpenLLM as a system service to create, start, and enable the application to infer the mosaicml/mpt-7b-instruct model as described in the steps below.

  1. View the openllm path

     $ which openllm



    Copy the OpenLLM path to your clipboard

  2. Create a new OpenLLM service file.

     $ sudo vim /etc/systemd/system/openllm.service
  3. Add the following configurations to the file. Replace the WorkingDirectory with the OpenLLM path (without openllm), and ExecStart with the actual OpenLLM path including the executable binary

     Description= Daemon for OpenLLM Demo Application
     ExecStart=/home/example_user/.local/bin/openllm start mpt --model-id "mosaicml/mpt-7b-instruct"

    Save and close the file

    The above configuration creates a new openllm system service used to manage OpenLLM. Replace User and Group with your actual non-root sudo user account details.

  4. Enable the openllm system service

     $ sudo systemctl enable openllm
  5. Start the service

     $ sudo systemctl start openllm
  6. Verify the service status and confirm if it's active and running

     $ sudo systemctl status openllm

    Your output should appear like the one below:

     ● openllm.service - Daemon for OpenLLM Demo Application
          Loaded: loaded (/etc/systemd/system/openllm.service; enabled; vendor preset: enabled)
          Active: active (running) since Sun 2023-08-27 11:32:07 UTC; 3s ago
        Main PID: 12077 (openllm)
           Tasks: 3 (limit: 72185)
          Memory: 251.4M
             CPU: 2.621s
          CGroup: /system.slice/openllm.service
                  - 12077 /usr/bin/python3 /home/example_user/.local/bin/openllm start mpt --model-id mosaicml/mpt-7b-instruct

Configure Nginx as a Reverse Proxy

To accept connections to the backend OpenLLM service running on port 3000, configure Nginx to work as a reverse proxy to handle HTTP connections and forward them to the backend port as described in the steps below.

  1. Verify that the Nginx web server is active and running

     $ sudo systemctl status nginx
  2. Create a new Nginx virtual host configuration file in the sites-available directory

     $ sudo vim /etc/nginx/sites-available/openllm
  3. Add the following configurations to the file. Replace with your actual subdomain.

     server {
         listen 80;
         listen [::]:80;
         location / {

    Save and close the file

  4. Link the configuration file to the sites-enabled directory to activate the configuration file

     $ sudo ln -s /etc/nginx/sites-available/openllm /etc/nginx/sites-enabled/
  5. Test the Nginx configuration for errors

     $ sudo nginx -t

    When your configuration is valid, the test should be successful as below

     nginx: the configuration file /etc/nginx/nginx.conf syntax is ok
     nginx: configuration file /etc/nginx/nginx.conf test is successful
  6. Restart Nginx to apply changes

     $ sudo systemctl reload nginx
  7. By default, UFW blocks all incoming connections, allow the HTTP port 80 through the firewall.

     $ sudo ufw allow 80/tcp
  8. Allow the HTTPS port 443 through the firewall

     $ sudo ufw allow 443/tcp
  9. Reload firewall rules to save changes

     $ sudo ufw reload

Secure OpenLLM with Valid Let's Encrypt SSL Certificates

To secure the OpenLLM interface in a production environment, generate valid Secure Socket Layer (SSL) certificates using a trusted CA such as Let's Encrypt. In this section, install the Certbot Let's Encrypt client to request SSL certificates for your domain name and bind them to your Nginx configuration as described in the steps below.

  1. Using snap, install the Certbot Let's Encrypt client

     $ sudo snap install --classic certbot
  2. Request a new SSL certificate for your domain. Replace with your actual domain name

     $ sudo certbot --nginx -d 

    When prompted, enter your email address and press Y to accept the Let's Encrypt terms.

  3. Test that the SSL certificate auto renews upon expiry

     $ sudo certbot renew --dry-run

    Let's Encrypt Certificates are valid for 90 days. If the above command does not throw any error, then Certbot handles auto renewal process is working correctly.

  4. Using a web browser such as Chrome, visit your domain to verify that the server accepts HTTPS connections

    When successful, the OpenLLM web user interface should load correctly in your browser.

    OpenLLM dashboard image

Access the OpenLLM Service

  1. Within the main OpenLLM service interface, find the service APIs section and click POST/v1/generate

    Post request button

  2. Verify the example value, and click the Try it out button

    Open LLM try it out button

  3. In the prompt value field, enter the prompt What are large language models

  4. Click Execute to run the prompt with example values

    Execute button

    You can manipulate the intensity of the response by changing the values of the listed parameters. Below is what each parameter does:

    • top_p: Chooses the best probability tokens for the output making the output more focused and relevant
    • epsilon_cutoff: Ignores the tokens with low probability, it cuts off tokens with probabilities lower than the epsilon value
    • diversity_penalty: Influences the diversity of the output. A higher parameter value creates a more diverse and less repetitive response
    • repition_penalty: Poses a penalty on the tokens that repeat consecutively in the generated output
    • length_penalty: Controls the length of the response, a higher parameter value generates a longer response and vice-versa
    • no_repeat_ngram_size: Penalizes the tokens forming n-grams(sequence of n tokens) that already appear in the response
    • remove_invalid_values: Automatically removes tokens with invalid values from the generated response
    • num_return_sequences: Controls the different number of sequences a model should generate in a response
    • frequency_penalty: Manipulates the frequency of certain tokens a model selects when generating the response
    • use_beam_search: Finds relevant continuations for response generation using beam search when the parameter value is true
    • ignore_eos: Ignores the end of sentence tokens during response generation when the parameter value is true
    • n: Represents the number of tokens in each generated response

    When the code execution is successful, a generated response with the 200 success code displays in your interface.


Response Generation using the Python Client Tool

OpenLLM provides the functionality of a built-in Python client which allows users to interact with the model in a terminal session. Test the response generation as described below.

  1. Create a new Python file named

     $ vim
  2. Add the following code to the file

     import openllm
     client = openllm.client.HTTPClient('http://localhost:3000')
     response = client.query('What is a Large Language Model')

    Save and close the file

  3. Run the Python application

     $ python3


     What is a Large Language Model (LLM)?
     Large language models (LLMs) are a type of artificial neural network that are trained on large amounts of text data.

    The application imports the openllm library and creates an HTTPClient instance. Then, it calls the client.query() function to make a query to the model. The generated tokens output to the response variable and the generated text displays in the terminal session.

Query OpenLLM

To generate a response you can also query the model in the terminal itself without using the web UI or the Python client. Query OpenLLM to generate a response as described in the steps below.

  1. Run the query in the terminal

     $ export OPENLLM_ENDPOINT=http://localhost:3000
     openllm query 'What is a large language model'

    The above code sets the environment variable OPENLLM_ENDPOINT. Then, a query forwards to the OpenLLM service using the openllm tool .

    Your command output should look like the one below:

     What is a large language model
     What is a large language model (LLM)?
     Large language models (LLMs) are a type of artificial intelligence (AI) that can read and understand text. They are trained on large amounts of text, such as books, news articles, and other written content.

Use OpenLLM with Langchain

OpenLLM provides many functionalities to use the potential of Large Language Models (LLM). Among these functionalities, OpenLLM allows building features on top of LLMs using OpenLLMs Langchain integration. In this section, use Langchain with OpenLLM to generate responses as described below.

  1. Create a new Python file named

     $ vim
  2. Add the following code to the file

     from langchain.llms import OpenLLM
     from langchain import PromptTemplate
     from langchain.chains import LLMChain
     from langchain.agents import load_tools
     from langchain.agents import initialize_agent
     from langchain.agents import AgentType
     from langchain.agents.load_tools import get_all_tool_names
     from langchain import ConversationChain
     llm = OpenLLM(model_name="mpt", model_id='mosaicml/mpt-7b-instruct')
     prompt = PromptTemplate(
         template="Here are the unique names of the companies that {product}:",
     chain = LLMChain(llm=llm, prompt=prompt)
     print("offer cloud services"))

    Save and close the file

    • The above application code imports the necessary modules from the langchain library, then it initializes the OpenLLM and mpt models
    • After the initialization is complete, a prompt template sets which is one of the features Langchain provides, an input_variable sets with an input variable product and a given template string, the purpose of the template is to generate output based on the product
    • The LLMChain initializes, this uses the model to generate responses based on the provided prompts
    • The method returns the actual input that goes into the prompt template, the actual input can change based on your needs. The prompt template outputs a list of companies that offer cloud services.
  3. Run the Python application

     $ python3


     Instantiating an MPTForCausalLM model from /home/example_user/.cache/huggingface/modules/transformers_modules/925e0d80e50e77aaddaf9c3ced41ca4ea23a1025/
     You are using config.init_device='cuda', but you can also use config.init_device="meta" with Composer + FSDP for fast initialization.
     Loading checkpoint shards: 100%|===============================================================| 2/2 [00:07<00:00,  3.97s/it]
     Here are the unique names of the companies that offer cloud services: Amazon Web Services (AWS), Microsoft Azure, Google Cloud Platform (GCP), IBM Cloud, Oracle Cloud, Alibaba Cloud, Tencent Cloud, Huawei Cloud, and Baidu Cloud.

Additional OpenLLM Functionality

  • You can use HuggingFace functionality on top of OpenLLM, although it's only supported for the Starcoder model

  • You can quantize models with OpenLLMs support of bitsandbytes and GPTQ. For example:

      $ openllm start mpt --quantize int8
  • You can create embeddings with the OpenLLM embeddings endpoint for supported models

  • You can incorporate custom Large Language models into the OpenLLM ecosystem. Visit the Add a new model repository for more information.


In this article you run OpenLLM, installed the CUDA Toolkit and required libraries to run the platform on a Vultr Cloud GPU server. Additionally,you generated responses from the MPT model using the OpenLLM web dashboard, terminal, built-in Python client, and Langchain integration. For more information about OpenLLM, visit the project repository.

Next Steps

To implement more solutions on your Vultr Cloud GPU server, visit the following resources: