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
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): https://github.com/qwopqwop200/GPTQ-for-LLaMa
| [Up-to-date triton branch](https://github.com/qwopqwop200/GPTQ-for-LLaMa) | 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](https://github.com/qwopqwop200/GPTQ-for-LLaMa/tree/cuda) | 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.
>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.
You are going to need to have a C++ compiler installed into your system for the last command. On Linux, `sudo apt install build-essential` or equivalent is enough.
⚠️ The tokenizer files in the sources above may be outdated. Make sure to obtain the universal LLaMA tokenizer as described [here](https://github.com/oobabooga/text-generation-webui/blob/main/docs/LLaMA-model.md#option-1-pre-converted-weights).
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:
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.
To load a model quantized with AutoGPTQ in the web UI, you need to first manually install the AutoGPTQ library:
```
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](https://github.com/oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md#step-0-install-nvcc)).
### 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.