A gradio web UI for running Large Language Models like LLaMA, llama.cpp, GPT-J, Pythia, OPT, and GALACTICA.
Go to file
2023-01-11 01:17:20 -03:00
models Add the default folders 2023-01-06 02:38:09 -03:00
presets Add a new preset 2023-01-07 15:13:00 -03:00
convert-to-torch.py Remove T5 support 2023-01-10 23:41:35 -03:00
download-model.py Fix the download script 2023-01-07 16:49:21 -03:00
html_generator.py Add nice HTML output for all models 2023-01-11 01:10:11 -03:00
LICENSE Initial commit 2022-12-21 01:17:38 -03:00
README.md Add nice HTML output for all models 2023-01-11 01:10:11 -03:00
requirements.txt Add script to download models 2023-01-06 19:57:31 -03:00
server.py Sort models alphabetically 2023-01-11 01:17:20 -03:00
webui.png Update the screenshot 2023-01-07 18:29:32 -03:00

text-generation-webui

A gradio webui for running large language models locally. Supports gpt-j-6B, gpt-neox-20b, opt, galactica, and many others.

Its goal is to become the AUTOMATIC1111/stable-diffusion-webui of text generation.

webui screenshot

Features

  • Switch between different models using a dropdown menu.
  • Generate nice HTML output for GPT-4chan.
  • Generate Markdown output for GALACTICA, including LaTeX support.
  • Notebook mode that resembles OpenAI's playground.
  • Chat mode for conversation and role playing.
  • Load parameter presets from text files.
  • Load large models in 8-bit mode.
  • Split large models across your GPU(s) and CPU.
  • CPU mode.

Installation

Create a conda environment:

conda create -n textgen
conda activate textgen

Install the appropriate pytorch for your GPU. For NVIDIA GPUs, this should work:

conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

Install the requirements:

pip install -r requirements.txt

Downloading models

Models should be placed under models/model-name. For instance, models/gpt-j-6B for gpt-j-6B.

Hugging Face

Hugging Face is the main place to download models. These are some of my favorite:

The files that you need to download are the json, txt, and pytorch*.bin files. The remaining files are not necessary.

For your convenience, you can automatically download a model from HF using the script download-model.py. Its usage is very simple:

python download-model.py organization/model

For instance:

python download-model.py facebook/opt-1.3b

GPT-4chan

GPT-4chan has been shut down from Hugging Face, so you need to download it elsewhere. You have two options:

You also need to put GPT-J-6B's config.json file in the same folder: config.json

Converting to pytorch (optional)

The script convert-to-torch.py allows you to convert models to .pt format, which is about 10x faster to load:

python convert-to-torch.py models/model-name

The output model will be saved to torch-dumps/model-name.pt. When you load a new model, the webui first looks for this .pt file; if it is not found, it loads the model as usual from models/model-name.

Starting the webui

conda activate textgen
python server.py

Then browse to

http://localhost:7860/?__theme=dark

Optionally, you can use the following command-line flags:

-h, --help      show this help message and exit
--model MODEL   Name of the model to load by default.
--notebook      Launch the webui in notebook mode, where the output is written to the same text
                box as the input.
--chat          Launch the webui in chat mode.
--cpu           Use the CPU to generate text.
--auto-devices  Automatically split the model across the available GPU(s) and CPU.
--load-in-8bit  Load the model with 8-bit precision.
--no-listen     Make the webui unreachable from your local network.

Presets

Inference settings presets can be created under presets/ as text files. These files are detected automatically at startup.

System requirements

Check the wiki for some examples of VRAM and RAM usage in both GPU and CPU mode.

Contributing

Pull requests, suggestions and issue reports are welcome.

Other projects

Make sure to also check out the great work by KoboldAI. I have borrowed some of the presets listed on their wiki after performing a k-means clustering analysis to select the most relevant subsample.