This is the official implementation of the algorithm part of "FPQVAR: Floating Point Quantization for Visual Autoregressive Model with FPGA Hardware Co-design"

Paper link: https://arxiv.org/abs/2505.16335

1. Install pip packages: pip install -r requirements.txt

2. Install the cuda kernel for floating point quantization: pip install ./quant

1. Download pretrained VAR model from VAR

2. Download the reference ground truth npz file: 256x256 and 512x512

Please follow the evaluation process as below. We demonstrate the commands for ImageNet 256x256 FP4 quantization. Commands for other configurations can be found in run.sh

1. Quantize and generate images

python evaluate_fp_quant_transform_rotate.py --quant --w_bit 4 --a_bit 4 --weight_quant per_group --act_quant per_group --act_sym --activation_fp_quant --weight_fp_quant --act_fp_type fp_e2 --weight_fp_type fp_e2 --fc2_fp_type fp_e1m2_neg_e2m1_pos --rotate --block_rotate --transform

2. Pack the generated images into an npz file

python pack_figs.py --file_path path_to_the_folder_storing_generated_images

3. Evaluate with OpenAI's evaluation toolkit

python openai_evaluator.py path_to_the_folder_storing_reference_images path_to_the_folder_storing_generated_images

The code of GHT-Aware Learnable Transformation (GALT) can be found in learnable_transformation/. The learned smoothing factors for VAR-30 (256x256) and VAR-36 (512x512) are available in learnable_transformation/best_lambda_var30 and learnable_transformation/best_lambda_var36, respectively.

If our work assists your research, please give us a star ⭐ or cite us using:

@article{wei2025fpqvar,
  title={FPQVAR: Floating Point Quantization for Visual Autoregressive Model with FPGA Hardware Co-design},
  author={Wei, Renjie and Xu, Songqiang and Guo, Qingyu and Li, Meng},
  journal={arXiv preprint arXiv:2505.16335},
  year={2025}
}