MiniGPT-4

Simplified local setup of MiniGPT-4 running in an Anaconda environment. Fixes for various Windows OS issues are provided, as well as links to pre-prepared Vicuna weights.

I've also included a simple MiniGPT-4 server that you can run locally that will respond to API requests, along with an example client that demonstrates how to interact with it.

Requirements

You'll need an Nvidia GPU with at least 12GB of VRAM (24GB+ is preferred). These instructions were tested on a Windows 10 machine with an Nvidia 3080Ti GPU, but should work on Linux as well (not tested).

Setup

[1] Install Anaconda, open the root terminal, and create a new environment (and activate it):

conda create --name minigpt4 python=3.9
conda activate minigpt4

[2] Install a couple required Python packages:

conda install -c anaconda git urllib3

[3] Clone the official MiniGPT-4 repository and switch to its directory:

git clone https://github.com/Vision-CAIR/MiniGPT-4.git
cd MiniGPT-4

[4] Install requirements:

curl -L -o requirements.txt -C - "https://raw.githubusercontent.com/rbbrdckybk/MiniGPT-4/main/requirements.txt"
pip install -r requirements.txt

Note that if you get an error while installing pycocotools on Windows, you may need to install the Microsoft C++ Build Tools. See this issue for more information.

[5] Install PyTorch:

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

[6] Prepare your own weights by following the official instructions (step 2) here. This involves waiting for your access request to be approved, then downloading ~200GB of LLaMA weights, and then using the Vicuna toolset to prepare working weights.

-OR-

Simply reference these pre-prepared weights (credit to wangrongsheng) by opening minigpt4/configs/models/minigpt4.yaml & edit line 16:

if you have 24GB+ of VRAM:

llama_model: "wangrongsheng/MiniGPT-4-LLaMA"

if you have 12GB+ of VRAM:

llama_model: "wangrongsheng/MiniGPT-4-LLaMA-7B"

[7] Download and reference the pretrained MiniGPT-4 checkpoint (links provided by official repo here @ step 3):

if you have 24GB+ of VRAM, download this checkpoint.

if you have 12GB+ of VRAM, download this checkpoint.

Place the downloaded checkpoint file in your MiniGPT-4 directory, then open eval_configs/minigpt4_eval.yaml and modify line 11:

if you have 24GB+ of VRAM:

ckpt: 'pretrained_minigpt4.pth'

if you have 12GB+ of VRAM (feel free to correct the typo in the official filename):

ckpt: 'prerained_minigpt4_7b.pth'

[8] If you're on Windows, you'll need to run these commands to fix a known issue with bitsandbytes:

pip uninstall bitsandbytes
pip install git+https://github.com/Keith-Hon/bitsandbytes-windows.git

You'll also need to place this DLL into your [Anaconda root directory]\envs\textgen\lib\site-packages\bitsandbytes folder.

Skip this step entirely if you're on Linux!

[9] (optional) Download my simple API server & client implementation: I've removed gradio and set MiniGPT-4 up as a simple Flask server that you can run locally to handle API requests. I've also coded a simple client example so you can see how to interact with it.

pip install Flask
curl -L -o api-server.py -C - "https://raw.githubusercontent.com/rbbrdckybk/MiniGPT-4/main/api-server.py"
curl -L -o api-client-example.py -C - "https://raw.githubusercontent.com/rbbrdckybk/MiniGPT-4/main/api-client-example.py"
mkdir img
curl -L -o img/simpsons.jpg -C - "https://raw.githubusercontent.com/rbbrdckybk/MiniGPT-4/main/img/simpsons.jpg"

See below for usage instructions!

Usage

Run the official gradio demo to verify that everything works:

python demo.py --cfg-path eval_configs/minigpt4_eval.yaml --gpu-id 0

Note that several large files (~15GB total) will be downloaded on the first run.

If you downloaded my API server & client in step 9 (verify that the official gradio demo works properly before continuing!), you can test them by starting the server with:

python api-server.py

Once the server is running, you can start the client with:

python api-client-example.py

You should see the client send the example image (img/simpsons.jpg) to the server and ask MiniGPT-4 several questions about it. Import the MiniGPT4_Client class from api-client-example.py into your own projects to easily interact with MiniGPT-4!