transformers-gptq-quant

A Python-based utility for quantizing GPT models using the Hugging Face 'transformers' library.

Description

This repository provides a Python script to quantize models with the Hugging Face 'transformers' and AutoGPTQ for 4 or 8 bit. Quantization can reduce the memory requirements of models, which may be beneficial for deployment on resource-constrained devices.

Features

• Supports quantizing various GPTQ precisions (8bit and 4bit).
• Allows for command-line or direct script useage.
• Compatible with a range of models and datasets.

Prerequisites

To utilize this utility, ensure that you've installed the dependencies listed in the requirements.txt file. You can install these using pip:

pip install -r requirements.txt

Usage

You can run the script in two ways:

1. By setting hardcoded values for the parameters at the top of the script under 'Hardcoded default values'.
2. By providing command-line arguments to override the hardcoded values.

Parameters:

• model_id: The huggingface repository or local directory containing a HF model. Example: 'teknium/openhermes', './hermes2', etc. It can be a folder or a huggingface directory.
• bits: The number of bits to use for quantization. Options: 4 and 8 - Default: 4
• dataset: The name of the dataset to use. Example: 'wikitext2', 'c4', etc. - Default: wikitext2
• group_size: The size of the group for quantization. Usually a power of 2. - Default: 128
• device_map: Device mapping configuration for loading the model. Example: 'auto', 'cpu', 'cuda:0', etc. - Default "auto"

Running the script

1. Using hardcoded values:

   • Edit the values of MODEL_ID, BITS, DATASET, GROUP_SIZE, and DEVICE_MAP at the top of the script.
   • Run the script without additional command-line arguments.

2. Using command-line arguments:

   • Use the following syntax:

   python quantize.py --model_id 'teknium/openhermes' --bits 8 --dataset 'wikitext2' --group_size 32 --device_map 'auto'

   Note: Command-line arguments will take precedence over hardcoded values.

Requirements

Ensure you have the necessary dependencies by checking the requirements.txt in this repo. Install them via:

pip install -r requirements.txt

Contributing

Feel free to contribute to this project by creating issues or submitting pull requests. Ensure that your contributions are in line with the project's aim and structure.

License

This project is open source, under the MIT license.