LOGART: PUSHING THE LIMIT OF EFFICIENT LOGARITHMIC POST-TRAINING QUANTIZATION

This repository contains the official PyTorch implementation for the ICLR 2026 paper "LogART: Pushing The Limit of Efficient Logarithmic Post-training Quantization".

Get Started

1. Clone this repository:

   git clone https://github.com/logart-lab/logart.git

   Then navigate to the desired directory:

   cd LogART/LogART-LLM

   Or:

   cd LogART/LogART-CNN

   Or:

   cd LogART/LogART-ViT

2. Install PyTorch:

   Ensure you have PyTorch installed. You can install PyTorch 1.10.0 with the following command:

   pip install torch==1.10.0 torchvision --index-url https://download.pytorch.org/whl/cu113

Pretrained Models

LogART-LLM

Pretrained models for LogART-LLM can be obtained using the transformers library (version 4.43.2). Ensure transformers is installed:

pip install transformers==4.43.2

For dataset loading and preprocessing, this project uses the datasets library, tested on version 2.20.0. Install it with:

pip install datasets==2.20.0

LogART-CNN

The pretrained models for LogART-CNN are sourced from BRECQ and can be accessed via torch.hub. For example, to load a pretrained ResNet-18 model:

res18 = torch.hub.load('yhhhli/BRECQ', model='resnet18', pretrained=True)

If you encounter a URLError during download (likely due to network issues), you can manually download the model using wget and place it in the appropriate directory:

wget https://github.com/yhhhli/BRECQ/releases/download/v1.0/resnet50_imagenet.pth.tar
mv resnet50_imagenet.pth.tar ~/.cache/torch/checkpoints

The load_state_dict_from_url function will check the ~/.cache/torch/checkpoints directory before attempting to download.

LogART-ViT

Pretrained models for LogART-ViT can be obtained using the timm library (version 0.9.2). Ensure timm is installed:

pip install timm==0.9.2

Alternatively, you can directly download checkpoints provided by AdaLog. For example:

wget https://github.com/GoatWu/AdaLog/releases/download/v1.0/deit_tiny_patch16_224.bin
mkdir -p ./checkpoint/vit_raw/
mv deit_tiny_patch16_224.bin ./checkpoint/vit_raw/

Usage

LogART-LLM

The code for LogART-LLM was modified based on aespa. To quantize and evaluate the LLM, use the following command:

python main.py --model_path facebook/opt-125m --calib_data c4 --nsamples 32 --seqlen 2048 --seed 0 --w_bits 3 --num_iters 500 --lr 0.05 --w_method log_dynamic -scale_search --rrweight 1 [--hardware_approx] --learn_rounding [--test_with_hardware_approx]
    

Command-Line Arguments

• --model_path: Path to the LLM model. Choices: facebook/opt-125m, facebook/opt-1.3b, facebook/opt-6.7b, meta-llama/Llama-2-7b, meta-llama/Llama-3.1-8B.
• --calib_data: Name of the calibration dataset used for quantization.
• --nsamples: Number of samples selected from the calibration dataset for quantization.
• --seqlen: Sequence length of the input text processed by the model.
• --seed: Random seed for ensuring reproducibility of experiments.
• --w_bits: Bitwidth for weight quantization.
• --num_iters: Number of iterations in learnable rounding reconstruction.
• --lr: Adaptive rounding learning rate.
• --w_method: Quantization method. Choices: log_2, log_sqrt2, log_dynamic.
• --scale_search: Whether to search for the best scaling factor.
• --rrweight: Weight of rounding cost vs the reconstruction loss.
• --hardware_approx: Apply 1+1/2 hardware approximation method to $\sqrt{2}$.
• --learn_rounding: Whether to learn weight-rounding policy based on AdaRound.
• --test_with_hardware_approx: Whether to test the quantized model with hardware approximation.

LogART-CNN

The code for LogART-CNN was modified based on BRECQ. To quantize and evaluate a single CNN model, use the following command:

python main.py --data_path DATA_DIR --arch resnet18 -seed 0 --n_bits_w 4 --channel_wise --scale_method log_dynamic --search_method layer_wise --search_samples 32 [--test_before_calibration] --iters_w 2000 --calib_lr 0.05 --weight 1 [--hardware_approx] [--test_with_hardware_approx]
 

Command-Line Arguments

• --data_path: Path to ImageNet data.
• --arch: Model architecture to use. Choices: resnet18, resnet50, mobilenetv2.
• --seed: Random seed for ensuring reproducibility of experiments.
• --n_bits_w: Bitwidth for weight quantization.
• --channel_wise: Enable channel-wise quantization for weights.
• --scale_method: Linear quantization method or logarithmic base. Choices: log_2, log_sqrt2, log_dynamic.
• --search_method: Dynamic base search method of the hyperparameter searching process for dynamic logarithmic quantization. Choices: 'tensor_wise', 'layer_wise'.
• --search_samples: Size of the hyperparameter searching dataset.
• --test_before_calibration: Test the quantization accuracy after hyperparameter searching and before reconstruction.
• --iters_w: Number of iterations in learnable rounding reconstruction.
• --calib_lr: Adaptive rounding learning rate.
• --weight: Weight of rounding cost vs the reconstruction loss.
• --hardware_approx: Apply hardware-level log $\sqrt{2}$ approximation to 1+1/2.
• --test_with_hardware_approx: Whether to test the quantized model with hardware approximation.

LogART-ViT

The code for LogART-ViT was modified based on APHQ. To quantize and evaluate a single ViT model, use the following command:

python main.py --dataset DATA_DIR --model vit_base -seed 0 --config ./configs/4bit/best.py --iters_w 2000 --lr 0.05 --scale-method log_dynamic -search_method block_wise --rrweight 1 [--hardware_approx] [--optimize] [--test_with_hardware_approx]

Command-Line Arguments

• --dataset: Path to the dataset.
• --model: Model architecture. Choices: vit_tiny, vit_small, vit_base, vit_large, deit_tiny, deit_small, deit_base.
• --seed: Random seed for ensuring reproducibility of experiments.
• --config: File path to import the Config class.
• --optimize: Perform learnable rounding reconstruction to the model.
• --iters_w: Number of iterations in learnable rounding reconstruction.
• --lr: Adaptive rounding learning rate.
• --scale-method: Quantization method. Choices: log_2, log_sqrt2, log_dynamic.
• --search_method: Dynamic base search method of the hyperparameter searching process for dynamic logarithmic quantization. Choices: 'tensor_wise', 'block_wise'.
• --rrweight: Weight of rounding cost vs the reconstruction loss.
• --hardware_approx: Apply 1+1/2 hardware approximation method to $\sqrt{2}$.
• --test_with_hardware_approx: Whether to test the quantized model with hardware approximation.

Results

Results will be stored in ./results.csv. The ablation results of LogART's key components on LLMs with 3-bit channel-wise weight quantization are shown in the table below:

DBS	SFS	ABS	LLR	Calib. Data	OPT-125M			LLaMA2-7B
					PPL	Time	Memory	PPL	Time	Memory
×	×	×	×	-	170.64	0.7 s	0.40 GB	60.16	13.0 s	9.8 GB
×	×	×	✓	32	38.55	61.3 s	0.75 GB	9.74	58.6 min	20.9 GB
×	×	✓	×	-	79.70	0.7 s	0.40 GB	8.28	13.2 s	9.8 GB
×	×	✓	✓	32	36.39	61.3 s	0.75 GB	6.44	58.2 min	20.9 GB
×	✓	×	×	32	38.41	12.1 s	0.75 GB	6.66	6.6 min	20.9 GB
×	✓	×	✓	32	33.21	64.6 s	0.75 GB	6.24	63.1 min	20.9 GB
×	✓	✓	×	32	35.15	12.2 s	0.75 GB	6.55	6.6 min	20.9 GB
×	✓	✓	✓	32	32.55	64.6 s	0.75 GB	6.23	63.5 min	20.9 GB
✓	×	×	×	32	66.63	3.8 s	0.75 GB	18.49	83.2 s	20.9 GB
✓	×	×	✓	32	35.46	62.7 s	0.75 GB	9.26	59.0 min	20.9 GB
✓	×	✓	×	32	47.92	3.8 s	0.75 GB	7.82	82.2 s	20.9 GB
✓	×	✓	✓	32	33.68	62.9 s	0.75 GB	6.38	59.1 min	20.9 GB
✓	✓	×	×	32	36.10	16.8 s	0.75 GB	6.56	17.9 min	20.9 GB
✓	✓	×	✓	32	32.37	75.0 s	0.75 GB	6.19	73.7 min	20.9 GB
✓	✓	✓	×	32	34.29	17.0 s	0.75 GB	6.45	17.9 min	20.9 GB
✓	✓	✓	✓	32	31.15	75.1 s	0.75 GB	6.14	74.2 min	20.9 GB

Citation

If you find this work is useful for your research, please cite our paper:

@inproceedings{xu2026logart,
title={Log{ART}: Pushing the Limit of Efficient Logarithmic Post-Training Quantization},
author={Xu, Jiawei and Zheng, Yi and Sun, Chenghe and Zhou, Taiyu and Zhang, Zuqi and Li, Jie and Zheng, Lirong and Zou, Zhuo},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=V85HbymBLW}
}