InfoQuant

June 3, 2026 ยท View on GitHub

Official implementation for InfoQuant: Reducing Quantization Information Error via Peak Suppression Orthogonal Transformation.

Paper: https://arxiv.org/abs/2605.26175

InfoQuant optimizes peak-suppression orthogonal transformation matrices for LLaMA-family models, then evaluates quantized models with optional Learning Activation Clipping (LAC).

Environment

Requirements:

  • Python >= 3.9
  • PyTorch >= 2.0
  • CUDA-capable GPU for actual model rotation, quantization, and evaluation

Install Python dependencies:

pip install -r requirements.txt

Install Fast Hadamard Transform:

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

Download LLaMA-2 or LLaMA-3 weights locally from Hugging Face before running experiments.

Run PSOT

sh scripts/get_matrix.sh <MODEL_PATH> <BLOCK_NUM> <W_BITS> <A_BITS> <KV_BITS>

Typical block settings:

  • LLaMA-2 7B, LLaMA-3 8B, LLaMA-2 13B: block_diag=2
  • LLaMA-2 70B, LLaMA-3 70B: block_diag=4

Run LAC Evaluation

sh scripts/eval_lac.sh <MODEL_PATH> <ROTATION_PATH> <USE_BF16> <CALI_BS> <W_BITS> <A_BITS> <KV_BITS>

For 70B models, reduce calibration batch size if GPU memory is tight. The original experiments used about 24 GB GPU memory, but CUDA/runtime overhead can vary.

Results

Perplexity on WikiText2 and average accuracy across nine zero-shot tasks. All models are evaluated using uniform quantization with InfoQuant.

#Bits (W-A-KV)MethodLLaMA-3 8BWikiLLaMA-2 7BWikiLLaMA-2 13BWikiLLaMA-2 70BWikiLLaMA-3 70BWiki
16-16-16FP1668.096.1465.215.4767.614.8871.593.3273.812.86
4-16-16InfoQuant67.366.4864.345.6067.274.9971.253.4073.253.50
4-4-16InfoQuant65.747.0762.845.8666.715.1570.823.6270.715.24
4-4-4InfoQuant65.577.1663.165.8966.335.1870.353.6470.215.39

Baseline results for RTN, SmoothQuant, GPTQ, AWQ, QuaRot, SpinQuant, and OSTQuant are referenced from:

OSTQuant: Refining Large Language Model Quantization with Orthogonal and Scaling Transformations for Better Distribution Fitting Xing Hu, Yuan Cheng, Dawei Yang, Zhixuan Chen, Zukang Xu, Jiangyong Yu, XUCHEN, Zhihang Yuan, Zhe Jiang, Sifan Zhou The Thirteenth International Conference on Learning Representations (ICLR), 2025 https://openreview.net/forum?id=rAcgDBdKnP