vLLM-Intercept

vLLM-Intercept is a Python library and server that allows you to seamlessly port your existing OpenAI API-based applications to use a local vLLM server running models like facebook/opt-125m. This enables you to run language models locally, reducing latency and improving privacy, without extensively modifying your existing codebase.

Table of Contents

• Features
• Prerequisites
• Installation
• Getting Started
  • Setting Up the vLLM Server
  • Using vllm_intercept in Your Application
• Usage Example
• Directory Structure
  • Source Code Files
    • vllm_intercept.py
    • daemon_server.py
• Contributing
• License

Features

• Seamless Integration: Intercept and redirect OpenAI API calls to your local vLLM server with minimal code changes.
• FastAPI Server Wrapper: A daemon server that ensures the vLLM server is running and translates API requests.
• Model Flexibility: Run any Hugging Face-compatible model locally with vLLM.
• Portability: Keep your existing codebase largely unmodified by simply importing the vllm_intercept module.

Prerequisites

• Python 3.8 or higher
• vLLM installed
• Required Python packages:
  • fastapi
  • uvicorn
  • grpcio
  • grpcio-tools
  • pydantic
  • openai

Installation

Clone the Repository

git clone --recursive-submodules https://github.com/your_username/vLLM-Intercept.git
cd vLLM-Intercept

Run the Install Script

Make the install script executable and run it:

chmod +x install.sh
./install.sh

This script will:

Install the required Python dependencies. Generate the gRPC code from generation_service.proto. Set up the daemon_server.py as a systemd service named vllm_intercept. Start the service and enable it to run on system startup. Note: The install script must be run with appropriate permissions to create systemd service files.

Getting Started

Using vllm_intercept in Your Application Copy vllm_intercept.py to Your Project

Ensure vllm_intercept.py is in your project directory or accessible via your Python path.

Modify Your Application

In your Python script, replace:

import openai

With:

import vllm_intercept  # Import the interceptor
import openai

Note: Ensure that import vllm_intercept comes before import openai.

Run Your Application

Run your script as you normally would. The vllm_intercept module will redirect API calls to your local vLLM server.

Usage Example

Suppose you have the following script using OpenAI's API:

import openai

openai.api_key = "your-api-key-here"

def ask_openai(question):
    response = openai.ChatCompletion.create(
        model="gpt-3.5-turbo",
        messages=[
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": question}
        ]
    )
    answer = response['choices'][0]['message']['content']
    return answer

if __name__ == "__main__":
    user_question = input("Enter your question: ")
    print("Thinking...")
    answer = ask_openai(user_question)
    print(f"Assistant: {answer}")

To use vLLM-Intercept, modify the script as follows:

import vllm_intercept  # Import the interceptor
import openai

def ask_openai(question):
    response = openai.ChatCompletion.create(
        model="gpt-3.5-turbo",  # This will be mapped to "facebook/opt-125m"
        messages=[
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": question}
        ]
    )
    answer = response['choices'][0]['message']['content']
    return answer

if __name__ == "__main__":
    user_question = input("Enter your question: ")
    print("Thinking...")
    answer = ask_openai(user_question)
    print(f"Assistant: {answer}")

That's it! The rest of your code remains unchanged.

Directory Structure Your project directory should look like this:

vLLM-Intercept/
├── README.md
├── requirements.txt
├── install.sh
├── vllm_intercept.py
├── daemon_server.py
├── generation_service.proto
├── generation_service_pb2.py
├── generation_service_pb2_grpc.py
└── your_application.py

README.md: Project documentation. requirements.txt: List of Python dependencies. install.sh: Install script that sets up the systemd service. vllm_intercept.py: Interceptor module to redirect API calls. daemon_server.py: Daemon server script. generation_service.proto: Protobuf definition file. generation_service_pb2.py and generation_service_pb2_grpc.py: Generated gRPC code. your_application.py: Your existing script modified to use vllm_intercept. Source Code Files vllm_intercept.py Path: /vLLM-Intercept/vllm_intercept.py

daemon_server.py Path: /vLLM-Intercept/daemon_server.py

install.sh Path: /vLLM-Intercept/install.sh

Contributing

Feel free to open an issue or create a pull request if you find a bug or want to add a new feature. Contributions are welcome!

License

This project is licensed under the GPLv3 License.

Additional Notes

Ensure All Dependencies Are Installed

The install script (install.sh) installs all required packages listed in requirements.txt.

Starting the Daemon Server

The install script sets up daemon_server.py as a systemd service named vllm_intercept. It starts automatically on system boot and can be managed using systemctl commands:

Check Service Status

sudo systemctl status vllm_intercept

Restart the Service

sudo systemctl restart vllm_intercept

Stop the Service

sudo systemctl stop vllm_intercept

Logs

The service logs are stored in the working directory:

Standard Output: vllm_intercept.log Standard Error: vllm_intercept_error.log

Testing the Setup

After setting up, test your application to ensure that it's correctly communicating with the local vLLM server.