We demonstrate that MXINT8/NVINT4 quantization formats are more accurate than their FP counterparts form both theoretical and practical scenarios. More details can be found in our paper INT v.s. FP: A Comprehensive Study of Fine-Grained Low-bit Quantization Formats.
- Theoretical QSNR
- Tensor-wise QSNR
cal_qsnr.py: calculate the QSNR of quantized models.cal_kl_ppl.py: calculate the KL divergence and perplexity of quantized models.quant/quant_func.py: fake quantization function of integer and float-point quantizationquant/quant_linear.py: linear layer with integer and float-point quantization
We offer the scripts to evaluate the KL divergence and perplexity of quantized models across different low-bit formats. You can change --model_path to evaluate the KL divergence and perplexity of other models. You can also set --rotate_dim 32 to introduce 32x32 random Hadamard rotation before quantization.
- MXINT8
CUDA_VISIBLE_DEVICES=0 python3 cal_kl_ppl.py --model_path ./hf_models/Llama-3.2-1B --q1_x 8 --q1_w 8 --group_size 32 --e8_scale --e8_scale_op ceil
- MXFP8
CUDA_VISIBLE_DEVICES=0 python3 cal_kl_ppl.py --model_path ./hf_models/Llama-3.2-1B --quant_type fp --e_bit 4 --m_bit 3 --q1_x 8 --q1_w 8 --group_size 32 --e8_scale --e8_scale_op ceil
- MXINT6
CUDA_VISIBLE_DEVICES=0 python3 cal_kl_ppl.py --model_path ./hf_models/Llama-3.2-1B --q1_x 6 --q1_w 6 --group_size 32 --e8_scale --e8_scale_op ceil
- MXFP6
CUDA_VISIBLE_DEVICES=0 python3 cal_kl_ppl.py --model_path ./hf_models/Llama-3.2-1B --quant_type fp --e_bit 2 --m_bit 3 --q1_x 6 --q1_w 6 --group_size 32 --e8_scale --e8_scale_op ceil
- MXINT4
CUDA_VISIBLE_DEVICES=0 python3 cal_kl_ppl.py --model_path ./hf_models/Llama-3.2-1B --q1_x 4 --q1_w 4 --group_size 32 --e8_scale --e8_scale_op ceil
- MXFP4
CUDA_VISIBLE_DEVICES=0 python3 cal_kl_ppl.py --model_path ./hf_models/Llama-3.2-1B --quant_type fp --e_bit 2 --m_bit 1 --q1_x 4 --q1_w 4 --group_size 32 --e8_scale --e8_scale_op ceil
- NVINT4
CUDA_VISIBLE_DEVICES=0 python3 cal_kl_ppl.py --model_path ./hf_models/Llama-3.2-1B --q1_x 4 --q1_w 4 --group_size 16 --scale_quant
- NVFP4
CUDA_VISIBLE_DEVICES=0 python3 cal_kl_ppl.py --model_path ./hf_models/Llama-3.2-1B --quant_type fp --e_bit 2 --m_bit 1 --q1_x 4 --q1_w 4 --group_size 16 --scale_quant
We offer the scripts to evaluate the QSNR of quantized models across different low-bit formats. You can change --model_path to evaluate the KL divergence and perplexity of other models. You can also set --rotate_dim 32 to introduce 32x32 random Hadamard rotation before quantization.
- MXINT8
CUDA_VISIBLE_DEVICES=0 python3 cal_qsnr.py \
--model_path ./hf_models/Llama-3.2-1B \
--bit 8 --e8_scale --e8_scale_op ceil \
--output_dir ./qsnr_results/llama3.2_1b/ \
--filename_prefix llama3.2_1b_base_mxint8
- MXFP8
CUDA_VISIBLE_DEVICES=0 python3 cal_qsnr.py \
--model_path ./hf_models/Llama-3.2-1B \
--bit 8 --e8_scale --e8_scale_op ceil \
--quant_type fp --e_bit 4 --m_bit 3 \
--output_dir ./qsnr_results/llama3.2_1b/ \
--filename_prefix llama3.2_1b_base_mxfp8
- MXINT6
CUDA_VISIBLE_DEVICES=0 python3 cal_qsnr.py \
--model_path ./hf_models/Llama-3.2-1B \
--bit 6 --e8_scale --e8_scale_op ceil \
--output_dir ./qsnr_results/llama3.2_1b/ \
--filename_prefix llama3.2_1b_base_mxint6
- MXFP6
CUDA_VISIBLE_DEVICES=0 python3 cal_qsnr.py \
--model_path ./hf_models/Llama-3.2-1B \
--bit 6 --e8_scale --e8_scale_op ceil \
--quant_type fp --e_bit 2 --m_bit 3 \
--output_dir ./qsnr_results/llama3.2_1b/ \
--filename_prefix llama3.2_1b_base_mxfp6
- MXINT4
CUDA_VISIBLE_DEVICES=1 python3 cal_qsnr.py \
--model_path ./hf_models/Llama-3.2-1B \
--bit 4 --e8_scale --e8_scale_op ceil \
--output_dir ./qsnr_results/llama3.2_1b/ \
--filename_prefix llama3.2_1b_base_mxint4
- MXFP4
CUDA_VISIBLE_DEVICES=2 python3 cal_qsnr.py \
--model_path ./hf_models/Llama-3.2-1B \
--bit 4 --e8_scale --e8_scale_op ceil \
--quant_type fp --e_bit 2 --m_bit 1 \
--output_dir ./qsnr_results/llama3.2_1b/ \
--filename_prefix llama3.2_1b_base_mxfp4
- NVINT4
CUDA_VISIBLE_DEVICES=0 python3 cal_qsnr.py \
--model_path ./hf_models/Llama-3.2-1B \
--bit 4 --scale_quant \
--output_dir ./qsnr_results/llama3.2_1b/ \
--filename_prefix llama3.2_1b_base_nvint4
- NVFP4
CUDA_VISIBLE_DEVICES=0 python3 cal_qsnr.py \
--model_path ./hf_models/Llama-3.2-1B \
--bit 4 --scale_quant \
--quant_type fp --e_bit 2 --m_bit 1 \
--output_dir ./qsnr_results/llama3.2_1b/ \
--filename_prefix llama3.2_1b_base_nvfp4
@article{int_vs_fp_2025,
title={INT v.s. FP: A Comprehensive Study of Fine-Grained Low-bit Quantization Formats},
author={Chen, Mengzhao and Wu, Meng and Jin, Hui and Yuan, Zhihang and Liu, Jing and Zhang, Chaoyi and Li, Yunshui and Huang, Jie and Ma, Jin and Xue, Zeyue and Liu, Zhiheng and Bin, Xingyan and Luo, Ping},
journal={arXiv preprint arXiv:2510.25602},
year={2025}
}