MCP Waifu Queue

This project implements an MCP (Model Context Protocol) server for a conversational AI "waifu" character, leveraging a text generation service with a Redis queue and GPU acceleration. It utilizes the FastMCP library for simplified server setup and management.

Table of Contents

• Features
• Architecture
• Prerequisites
• Installation
• Configuration
• Running the Service
• MCP API
• Testing
• Troubleshooting
• Contributing
• License

Features

• Text generation using the distilgpt2 language model.
• Request queuing using Redis for handling concurrent requests.
• GPU support for faster inference.
• MCP-compliant API using FastMCP.
• Job status tracking.

Architecture

The project consists of several key components:

• main.py: The main entry point, initializing the FastMCP application.
• respond.py: The core text generation service, loading the distilgpt2 model and generating responses.
• queue.py: Handles interactions with the Redis queue, enqueuing requests and managing job IDs.
• worker.py: A Redis worker that processes jobs from the queue, utilizing respond.py for text generation.
• config.py: Manages configuration via environment variables.
• models.py: Defines Pydantic models for request and response validation.

The flow of a request is as follows:

1. A client sends a request to the generate_text MCP tool (defined in main.py).
2. The tool enqueues the request to a Redis queue (handled by queue.py).
3. A worker.py process picks up the request from the queue.
4. The worker calls the call_predict_response function (in utils.py), which interacts with respond.py to generate the text.
5. The generated text is stored, and the job status is updated.
6. The client can retrieve the result using the get_job_status resource (defined in main.py).

Prerequisites

• Python 3.7+
• pip
• Redis server (installed and running)
• A CUDA-enabled GPU (optional, but recommended for performance)

You can find instructions for installing Redis on your system on the official Redis website: https://redis.io/docs/getting-started/

Installation

1. Clone the repository:

   git clone <YOUR_REPOSITORY_URL>
   cd mcp-waifu-queue
   

2. Create and activate a virtual environment:

   python3 -m venv venv
   source venv/bin/activate  # On Linux/macOS
   venv\Scripts\activate  # On Windows
   

3. Install dependencies:

    pip install --user -r requirements.txt #If requirements.txt exists
   

   Or, if using pyproject.toml

   pip install --user -e .
   

Configuration

1. Copy the .env.example file to .env:

   cp .env.example .env
   

2. Modify the .env file to set the appropriate values for your environment. The following environment variables are available:

   • MODEL_PATH: The path to the pre-trained language model (default: distilgpt2).
   • GPU_SERVICE_URL: The URL of the GPU service (default: http://localhost:5001). This is used internally by the worker.
   • REDIS_URL: The URL of the Redis server (default: redis://localhost:6379).
   • QUEUE_PORT: The port for the queue service (default: 5000). This is no longer directly used for external access, as we're using MCP.
   • RESPOND_PORT: The port for the response service (default: 5001). This is used internally by the worker.
   • MAX_NEW_TOKENS: The maximum number of new tokens to generate (default: 20).

   Note: The .env file should not be committed to the repository for security reasons.

Running the Service

Start the services using the scripts/start-services.sh script:

./scripts/start-services.sh

This script will start the Redis server (if not already running), the worker, the queue service, and the response service. The services will run in the background.

MCP API

The server provides the following MCP-compliant endpoints:

Tools

• generate_text (prompt: str): Sends a text generation request and returns a job ID.

Resources

• job://{job_id}: Retrieves the status of a job. The response will include a status field (e.g., "queued", "processing", "completed", "failed") and, if completed, a result field containing the generated text.

Testing

The project includes tests. You can run all tests using pytest:

pytest tests

Troubleshooting

• Error: "Missing 'prompt' parameter": Make sure you are sending a prompt string to the generate_text tool.
• Error: "Error calling GPU service": Ensure that the respond.py service is running and accessible at the configured GPU_SERVICE_URL.
• Error: "Service unavailable": Check if Redis server, worker, queue and respond services are running.
• If encountering CUDA errors: Ensure your CUDA drivers and toolkit are correctly installed and compatible with your PyTorch version.

Contributing

1. Fork the repository.
2. Create a new branch for your feature or bug fix.
3. Commit your changes.
4. Push your branch to your forked repository.
5. Create a pull request.

Please adhere to the project's code of conduct.

License

This project is licensed under the MIT-0 License - see the LICENSE file for details.