Neural Compressor

An open-source Python library supporting popular model compression techniques for ONNX

---

Neural Compressor aims to provide popular model compression techniques inherited from Intel Neural Compressor yet focused on ONNX model quantization such as SmoothQuant, weight-only quantization through ONNX Runtime. In particular, the tool provides the key features, typical examples, and open collaborations as below:

• Support a wide range of Intel hardware such as Intel Xeon Scalable Processors, Intel Xeon CPU Max Series; support AMD CPU, ARM CPU, and NVidia GPU with limited testing

• Validate popular LLMs such as LLama2 and broad models such as BERT-base, and ResNet50 from popular model hubs such as Hugging Face, ONNX Model Zoo, by leveraging automatic accuracy-driven quantization strategies

• Collaborate with software platforms such as Microsoft Olive, and open AI ecosystem such as Hugging Face, ONNX and ONNX Runtime

Installation

Install from source

git clone https://github.com/onnx/neural-compressor.git
cd neural-compressor
pip install -r requirements.txt
pip install .

Note: Further installation methods can be found under Installation Guide.

Getting Started

Setting up the environment:

pip install onnx-neural-compressor "onnxruntime>=1.17.0" onnx

After successfully installing these packages, try your first quantization program.

Notes: please install from source before the formal pypi release.

Weight-Only Quantization (LLMs)

Following example code demonstrates Weight-Only Quantization on LLMs, it supports Intel CPU, Nvidia GPU, device will be selected for efficiency automatically when multiple devices are available.

Run the example:

from onnx_neural_compressor.quantization import matmul_nbits_quantizer

algo_config = matmul_nbits_quantizer.RTNWeightOnlyQuantConfig()
quant = matmul_nbits_quantizer.MatMulNBitsQuantizer(
    model,
    n_bits=4,
    block_size=32,
    is_symmetric=True,
    algo_config=algo_config,
)
quant.process()
best_model = quant.model

Static Quantization

from onnx_neural_compressor.quantization import quantize, StaticQuantConfig
from onnx_neural_compressor.quantization.calibrate import CalibrationDataReader


class DataReader(CalibrationDataReader):
    def __init__(self):
        self.encoded_list = []
        # append data into self.encoded_list

        self.iter_next = iter(self.encoded_list)

    def get_next(self):
        return next(self.iter_next, None)

    def rewind(self):
        self.iter_next = iter(self.encoded_list)


data_reader = DataReader()
config = StaticQuantConfig(calibration_data_reader=data_reader)
quantize(model, output_model_path, config)

Documentation

Overview
Architecture			Workflow			Examples
Feature
Quantization				SmoothQuant
Weight-Only Quantization (INT8/INT4)				Layer-Wise Quantization

Additional Content

• Contribution Guidelines
• Security Policy

Communication

• GitHub Issues: mainly for bug reports, new feature requests, question asking, etc.
• Email: welcome to raise any interesting research ideas on model compression techniques by email for collaborations.