<ul><li>Researchers introduce GPTQv2, a finetuning-free quantization method for compressing large-scale transformer architectures.</li><li>GPTQv2 uses asymmetric calibration to match the quantized layer's output to the output in the full-precision model, reducing quantization error.</li><li>Various techniques such as channel parallelization and Cholesky reformulation are utilized to parallelize the solution calculation in GPTQv2.</li><li>GPTQv2 demonstrates improved performance in low-bit quantization on a 405B language transformer and EVA-02 vision transformer.</li></ul>

GPTQv2: Efficient Finetuning-Free Quantization for Asymmetric Calibration

Discover more