Large Language Models Weight Compression Example

This example demonstrates how to optimize Large Language Models (LLMs) in ONNX format using NNCF weight compression API. The example applies 4/8-bit mixed-precision quantization to weights of Linear (Fully-connected) layers of TinyLlama/TinyLlama-1.1B-Chat-v1.0 model. This leads to a significant decrease in model footprint and performance improvement with OpenVINO Runtime.

Prerequisites

Before running this example, ensure you have Python 3.10+ installed and set up your environment:

1. Create and activate a virtual environment

python3 -m venv nncf_env
source nncf_env/bin/activate  # On Windows: nncf_env\Scripts\activate.bat

2. Install NNCF and other dependencies

python3 -m pip install ../../../../ -r requirements.txt

Run Example

To run example:

python main.py

This will automatically:

• Download the TinyLlama model and dataset
• Apply weight compression using NNCF
• Save the optimized model

Set ONNX Opset (Optional)

The exported model uses ONNX opset version 21 by default. You can override this by specifying a different opset version when running the script. For example:

python main.py 14