local-chatgpt-app

A complete LLM ChatBot UI and server, all running on your laptop (tested on macOS):

graph LR;
    subgraph laptop [Your Laptop Environment]
    ChatBot[ChatBot UI] -- HTTP --> LLM[LLM Web Server];
    end

• Runs a ChatGPT-like LLM (e.g. Llama 2) locally (using llama-cpp-python)
• Runs a ChatGPT-like UI/app locally (using chainlit)

Setup

This loosely follows the setup steps from https://llama-cpp-python.readthedocs.io/en/latest/install/macos/.

Python Version

This code has been tested on Python version 3.9.16. Pyenv is a great way to install Python.

Python Library Installation

Set up a virtual environment and install Python requirements:

python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

Model Download

Model download command:

huggingface-cli download \
    TheBloke/Llama-2-7B-Chat-GGUF \
    llama-2-7b-chat.Q5_K_M.gguf \
    --local-dir ./models/llama-7b-chat/ \
    --local-dir-use-symlinks False

This model file format (GGUF) is used for running LLM inference on macOS. To use a different model:

1. Update the above download command to use a different huggingface repo and filename (must be in GGUF format).
2. Update the run-server.sh script's MODEL variable with the path to the .gguf file after download.

Usage

Model Server

In one terminal, activate the virtualenv and run the model server with:

./run-server.sh

This will run the Llama 7B chat model inference server (by default on port 8000).

Chat UI

In another terminal, activate the virtualenv and run the UI with:

./run-ui.sh

This will open a ChatGPT-like UI in your browser using Chainlit (by default on port 8001), connected to the inference server.

Troubleshooting

Llama Library File Override

It may be necessary to run the llama_cpp.server with an override configured for LLAMA_CPP_LIB to the path of the .dylib file created in the environment by the llama-cpp-python package. This can be found by running find /path/to/venv -name '*.dylib' in your venv (or Conda env). Attempting to build the llama C++ from scratch using https://github.com/ggerganov/llama.cpp resulted in a segmentation fault for me when running the llama_cpp.server module, so the library override using a pre-built library file seems to be more stable.