DuQuant: Distributing Outliers via Dual Transformation Makes Stronger Quantized LLMs

Quantization of large language models (LLMs) faces significant challenges, particularly due to the presence of outlier activations that impede efficient low-bit representation.