uni-quant-cuda 0.2.4

Uni-Quant: CUDA-accelerated quantization/dequantization for TensorFlow models

Uni-Quant

Small library to quantize/dequantize TensorFlow models using PyTorch CUDA kernels.

Requirements

• Python: 3.13.13 (haven't tested on any other)
• CUDA Toolkit: >=12.8
• Python Dependencies: All required packages are listed in requirements.txt

Installing Dependencies

pip install -r requirements.txt

Installation from pip

pip install uni-quant-cuda

Usage

Importing Functions

For installed package (from pip):

from Uniquant import quantize, dequantize, dequantize_save

For local clone repo:

from uniquant import quantize, dequantize, dequantize_save

Main Functions

quantize(model_path, quant_directory="", quant_name="", pack_size=32, quant_size=4, overwrite=False)

Quantizes a TensorFlow or XGBoost model.

Arguments:

• model_path (str): Path to the model to quantize (with extension)
• quant_directory (str): Directory path to save the quantized model
• quant_name (str): Filename for the quantized model
• pack_size (int): Number of weights in one quantization batch (must be divisible by 2)
• quant_size (int): Number of bits per weight (available: 4 or 8)
• overwrite (bool): Whether to overwrite existing file

dequantize(quant_path, literal=False, balanced=True)

Dequantizes a model and returns it.

Arguments:

• quant_path (str): Path to the .uniq file to dequantize
• literal (bool): Whether weights should be unscaled
• balanced (bool): Whether weights should be balanced around 0

dequantize_save(quant_path, model_directory="", model_name="", overwrite=False)

Dequantizes a model, saves it, and returns it.

Arguments:

• quant_path (str): Path to the .uniq file to dequantize
• model_directory (str): Directory path to save the dequantized model
• model_name (str): Filename for the dequantized model
• overwrite (bool): Whether to overwrite existing file

Notes

• This package compiles CUDA kernels at runtime using torch.utils.cpp_extension.load_inline.
• Installing and using the CUDA compilation requires a compatible CUDA toolkit on the target machine (tested with >=12.8).