leanquant 0.1.1

The inference kernels for LeanQuant models.

LeanQuant

Overview

This package provides efficient inference kernels for running non-uniformly quantized LeanQuant models on CUDA-enabled GPUs. LeanQuant is a scalable and accurate quantization algorithm that compresses large language models by 4-8x while maintaining competitive performance.

Installation

Ensure your GPU supports CUDA 11 or CUDA 12. You can check your CUDA version with the command nvidia-smi | grep CUDA.

To install:

# For CUDA 11.x
pip install leanquant[cuda11]

# For CUDA 12.x
pip install leanquant[cuda12]

Models

Quantized LeanQuant models are available for download on our HuggingFace page: huggingface.co/LeanQuant

Technical Details

LeanQuant introduces an innovative loss-error-aware grid approach to quantization that significantly outperforms traditional methods. Our technique:

• Achieves superior compression ratios: Reduces model size by 4-8x without sacrificing capability
• Preserves model intelligence: Maintains performance comparable to full-precision models across challenging benchmarks
• Optimizes GPU execution: Features custom CUDA kernels specifically designed for non-uniform quantization format

The algorithm strategically allocates quantization precision based on parameter sensitivity, ensuring computational resources are focused where they matter most.

For a comprehensive explanation of our methodology and benchmark results, please refer to our research paper.

Citation

If you find LeanQuant useful in your research or applications, please consider citing our work:

@inproceedings{
    zhang2025leanquant,
    title={LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid},
    author={Tianyi Zhang and Anshumali Shrivastava},
    booktitle={The Thirteenth International Conference on Learning Representations},
    year={2025},
    url={https://openreview.net/forum?id=ISqx8giekS}
}