text-generation-webui/README.md
2023-01-09 18:14:09 -03:00

5.7 KiB

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 13b/20b models in 8-bit mode.
  • Load parameter presets from text files.
  • 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

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:

--model model-name: Load this model 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 instead of the GPU.

Presets

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

System requirements

These are the VRAM (in GiB) and RAM (in MiB) requirements to run some model examples.

GPU mode (default)

model VRAM (GPU) RAM
OPT-350M-Erebus 0.62 1939.3
arxiv_ai_gpt2 1.48 6350.68
blenderbot-1B-distill 2.38 2705.9
opt-1.3b 2.45 2868.12
gpt-neo-1.3b 2.54 4047.04
gpt4chan_model_float16 11.38 1909.79
gpt-j-6b-float16 11.38 2847.75
gpt-j-6B 11.38 3959.55
galactica-6.7b 12.4 1933.19
opt-6.7b 12.4 1944.21
bloomz-7b1-p3 13.17 1845.58

GPU mode with 8-bit precision

Allows you to load models that would not normally fit into your GPU. Enabled by default for 13b and 20b models in this webui.

model VRAM (GPU) RAM
OPT-13B-Erebus 12.23 749.08
opt-13b 12.23 1258.95
gpt-neox-20b 19.91 2104.04

CPU mode

A lot slower, but does not require a GPU.

model RAM
OPT-350M-Erebus 2622.17
arxiv_ai_gpt2 3764.81
gpt-neo-1.3b 5937.81
opt-1.3b 7346.08
blenderbot-1B-distill 7565.36
bloomz-7b1-p3 23613.9
gpt-j-6B 23975.5
gpt4chan_model 23999.5
gpt-j-6b-float16 24999.1
galactica-6.7b 26248
opt-6.7b 27334.2

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.