text-generation-webui/docs/GPTQ-models-(4-bit-mode).md
2023-07-17 12:51:01 -07:00

8.1 KiB

GPTQ is a clever quantization algorithm that lightly reoptimizes the weights during quantization so that the accuracy loss is compensated relative to a round-to-nearest quantization. See the paper for more details: https://arxiv.org/abs/2210.17323

4-bit GPTQ models reduce VRAM usage by about 75%. So LLaMA-7B fits into a 6GB GPU, and LLaMA-30B fits into a 24GB GPU.

Overview

There are two ways of loading GPTQ models in the web UI at the moment:

  • Using AutoGPTQ:

    • supports more models
    • standardized (no need to guess any parameter)
    • is a proper Python library
    • ~no wheels are presently available so it requires manual compilation~
    • supports loading both triton and cuda models
  • Using GPTQ-for-LLaMa directly:

    • faster CPU offloading
    • faster multi-GPU inference
    • supports loading LoRAs using a monkey patch
    • requires you to manually figure out the wbits/groupsize/model_type parameters for the model to be able to load it
    • supports either only cuda or only triton depending on the branch

For creating new quantizations, I recommend using AutoGPTQ: https://github.com/PanQiWei/AutoGPTQ

AutoGPTQ

Installation

No additional steps are necessary as AutoGPTQ is already in the requirements.txt for the webui. If you still want or need to install it manually for whatever reason, these are the commands:

conda activate textgen
git clone https://github.com/PanQiWei/AutoGPTQ.git && cd AutoGPTQ
pip install .

The last command requires nvcc to be installed (see the instructions above).

Usage

When you quantize a model using AutoGPTQ, a folder containing a filed called quantize_config.json will be generated. Place that folder inside your models/ folder and load it with the --autogptq flag:

python server.py --autogptq --model model_name

Alternatively, check the autogptq box in the "Model" tab of the UI before loading the model.

Offloading

In order to do CPU offloading or multi-gpu inference with AutoGPTQ, use the --gpu-memory flag. It is currently somewhat slower than offloading with the --pre_layer option in GPTQ-for-LLaMA.

For CPU offloading:

python server.py --autogptq --gpu-memory 3000MiB --model model_name

For multi-GPU inference:

python server.py --autogptq --gpu-memory 3000MiB 6000MiB --model model_name

Using LoRAs with AutoGPTQ

Not supported yet.

GPTQ-for-LLaMa

GPTQ-for-LLaMa is the original adaptation of GPTQ for the LLaMA model. It was made possible by @qwopqwop200: https://github.com/qwopqwop200/GPTQ-for-LLaMa

Different branches of GPTQ-for-LLaMa are currently available, including:

Branch Comment
Old CUDA branch (recommended) The fastest branch, works on Windows and Linux.
Up-to-date triton branch Slightly more precise than the old CUDA branch from 13b upwards, significantly more precise for 7b. 2x slower for small context size and only works on Linux.
Up-to-date CUDA branch As precise as the up-to-date triton branch, 10x slower than the old cuda branch for small context size.

Overall, I recommend using the old CUDA branch. It is included by default in the one-click-installer for this web UI.

Installation

Start by cloning GPTQ-for-LLaMa into your text-generation-webui/repositories folder:

mkdir repositories
cd repositories
git clone https://github.com/oobabooga/GPTQ-for-LLaMa.git -b cuda

If you want to you to use the up-to-date CUDA or triton branches instead of the old CUDA branch, use these commands:

git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa.git -b cuda
git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa.git -b triton

Next you need to install the CUDA extensions. You can do that either by installing the precompiled wheels, or by compiling the wheels yourself.

Precompiled wheels

Kindly provided by our friend jllllll: https://github.com/jllllll/GPTQ-for-LLaMa-Wheels

Windows:

pip install https://github.com/jllllll/GPTQ-for-LLaMa-Wheels/raw/main/quant_cuda-0.0.0-cp310-cp310-win_amd64.whl

Linux:

pip install https://github.com/jllllll/GPTQ-for-LLaMa-Wheels/raw/Linux-x64/quant_cuda-0.0.0-cp310-cp310-linux_x86_64.whl

Manual installation

Step 1: install nvcc

conda activate textgen
conda install -c conda-forge cudatoolkit-dev

The command above takes some 10 minutes to run and shows no progress bar or updates along the way.

You are also going to need to have a C++ compiler installed. On Linux, sudo apt install build-essential or equivalent is enough.

If you're using an older version of CUDA toolkit (e.g. 11.7) but the latest version of gcc and g++ (12.0+), you should downgrade with: conda install -c conda-forge gxx==11.3.0. Kernel compilation will fail otherwise.

Step 2: compile the CUDA extensions

cd repositories/GPTQ-for-LLaMa
python setup_cuda.py install

Getting pre-converted LLaMA weights

  • Direct download (recommended):

https://huggingface.co/Neko-Institute-of-Science/LLaMA-7B-4bit-128g

https://huggingface.co/Neko-Institute-of-Science/LLaMA-13B-4bit-128g

https://huggingface.co/Neko-Institute-of-Science/LLaMA-30B-4bit-128g

https://huggingface.co/Neko-Institute-of-Science/LLaMA-65B-4bit-128g

These models were converted with desc_act=True. They work just fine with ExLlama. For AutoGPTQ, they will only work on Linux with the triton option checked.

  • Torrent:

https://github.com/oobabooga/text-generation-webui/pull/530#issuecomment-1483891617

https://github.com/oobabooga/text-generation-webui/pull/530#issuecomment-1483941105

These models were converted with desc_act=False. As such, they are less accurate, but they work with AutoGPTQ on Windows. The 128g versions are better from 13b upwards, and worse for 7b. The tokenizer files in the torrents are outdated, in particular the files called tokenizer_config.json and special_tokens_map.json. Here you can find those files: https://huggingface.co/oobabooga/llama-tokenizer

Starting the web UI:

Use the --gptq-for-llama flag.

For the models converted without group-size:

python server.py --model llama-7b-4bit --gptq-for-llama 

For the models converted with group-size:

python server.py --model llama-13b-4bit-128g  --gptq-for-llama --wbits 4 --groupsize 128

The command-line flags --wbits and --groupsize are automatically detected based on the folder names in many cases.

CPU offloading

It is possible to offload part of the layers of the 4-bit model to the CPU with the --pre_layer flag. The higher the number after --pre_layer, the more layers will be allocated to the GPU.

With this command, I can run llama-7b with 4GB VRAM:

python server.py --model llama-7b-4bit --pre_layer 20

This is the performance:

Output generated in 123.79 seconds (1.61 tokens/s, 199 tokens)

You can also use multiple GPUs with pre_layer if using the oobabooga fork of GPTQ, eg --pre_layer 30 60 will load a LLaMA-30B model half onto your first GPU and half onto your second, or --pre_layer 20 40 will load 20 layers onto GPU-0, 20 layers onto GPU-1, and 20 layers offloaded to CPU.

Using LoRAs with GPTQ-for-LLaMa

This requires using a monkey patch that is supported by this web UI: https://github.com/johnsmith0031/alpaca_lora_4bit

To use it:

  1. Clone johnsmith0031/alpaca_lora_4bit into the repositories folder:
cd text-generation-webui/repositories
git clone https://github.com/johnsmith0031/alpaca_lora_4bit

⚠️ I have tested it with the following commit specifically: 2f704b93c961bf202937b10aac9322b092afdce0

  1. Install https://github.com/sterlind/GPTQ-for-LLaMa with this command:
pip install git+https://github.com/sterlind/GPTQ-for-LLaMa.git@lora_4bit
  1. Start the UI with the --monkey-patch flag:
python server.py --model llama-7b-4bit-128g --listen --lora tloen_alpaca-lora-7b --monkey-patch