Code for ``DartQuant: Efficient Rotational Distribution Calibration for LLM Quantization''

1. Requirements:

• python 3.10, pytorch >= 2.0

• install pytorch with cuda from https://pytorch.org/get-started/locally/, it is prerequisite for fast-hadamard-transform package.

• pip install -r requirement.txt

  install fast-hadamard-transform

  cd third-part
  git clone https://github.com/Dao-AILab/fast-hadamard-transform.git
  cd fast-hadamard-transform
  pip install .
  

  install lm-eval

  git clone https://github.com/EleutherAI/lm-evaluation-harness.git
  cd lm-evaluation-harness
  pip install -e .
  

Guidelines

• The ./fake_quant folder contains the code for fusing the calibrated rotation matrix and performing the quantization test. The usage is described in detail in the Readme.md file in the directory.

• The ./calibrater folder contains the code for obtaining the calibration set and the calibration rotation matrix. The specific usage is described in the Readme.md in this directory.

NPU

• ./NPU_DartQuant folder contains contains NPU runtime code, and its usage is basically the same as that of the GPU version.