Quantized Inference on Generative LLMs (QIGen)

Code generator for inference on Quantized Large Language Models. Quantization is done using GPTQ.

Current features

• Support for LlaMA and OPT
• 4,3, and 2 bit inference
• x86 with AVX2 support
• Support for pyTorch and transformers
• Support for generic quantization group size

TODOs

• Support for ARM Neon
• Support for AVX512
• Including quantization error analysis in code generation

Usage

Installation

1. Install dependencies via pip install -r requirements.txt
2. Install transformers from source pip install git+https://github.com/huggingface/transformers
3. Install the python module python setup.py install. This will run a search to find the best parameters for register usage.

Example

We give an example notebook in demo.ipynb. The basic workflow is

• load floating point model,
• load quantized checkpoint from GPTQ,
• call the infergen.swap_modules_llama(model, quantized_checkpoint, bits=4, p=64, l1=l1, inplace=False) function, where model is the full-size model, quantized_checkpoint is the quantized model, bits is the number of bits used for the quantization,l1 is the size of the l1 data cache in bits, p is the number of cores to use, and inplace is a flag to swap in place or creating a copy.
• Use the quantized model as a normal transformer.