RTN: Round to Nearest Quantization (Qwen3)

The quantization and evaluation of RTN are encapsulated in the GPTQ method. By following the same installation steps as GPTQ, you can run RTN quantization and perform evaluations.

Installation Setup

1. You can clone the GPTQ-for-LLaMA repository and Replace the qwen.py file in the GPTQ-for-Qwen directory.

git clone https://github.com/qwopqwop200/GPTQ-for-LLaMa
cd GPTQ-for-LLaMa

​ For more details about the GPTQ-for-LLaMA repository, please refer to this link.

​ OR You can simply clone our repository, the directory GPTQ-for-Qwen can be directly run and used.

2. Install required packages

conda create -n gptq python=3.10 -y
conda activate gptq
pip install --upgrade pip  # enable PEP 660 support

File Replacement

Before starting quantization, you need to replace some files to support the Qwen3 model:

• Add eval_my directory: Place the eval_my directory under the GPTQ-for-Qwen directory.

Quantization and Evaluation

Group-wise Quantization

1. Perform RTN search and save results

CUDA_VISIBLE_DEVICES=0 python path_of_qwen.py your_model_path c4(Validation dataset) \
--wbits 4 --true-sequential --act-order \
--groupsize 128  --model_name gptq_4B_w4_128 \
--save gptq_4B_w4_128.pth --nearest

2. Evaluate the quantized model

CUDA_VISIBLE_DEVICES=0 python path_of_qwen.py your_model_path c4(Validation dataset) \
--wbits 4 --true-sequential --act-order \
--groupsize 128  --load gptq_4B_w4_128.pth \
--eval --nearest

Per-channel Quantization

1. Perform RTN search and save results

CUDA_VISIBLE_DEVICES=0 python path_of_qwen.py your_model_path c4(Validation dataset) \
--wbits 4 --true-sequential --act-order \
--groupsize -1  --model_name gptq_4B_w4_128 \
--save gptq_4B_w4_128.pth --nearest

2. Evaluate the quantized model

CUDA_VISIBLE_DEVICES=0 python path_of_qwen.py your_model_path c4(Validation dataset) \
--wbits 4 --true-sequential --act-order \
--groupsize -1  --load gptq_4B_w4_128.pth \
--eval --nearest

Notes

• Available quantization bit-widths (w_bit): 2, 4, 8.
• Use the --groupsize parameter (e.g., 128) for group-wise quantization
• Set the groupsize parameter to -1 for per-channel quantization
• The model_name parameter is chosen as a suffix for storing the result tables of MMLU. It is not very important.
• Make sure you have sufficient GPU memory to run a 32B-sized model