A Comprehensive Evaluation on Quantization Techniques for Large Language Models

Abstract: For LLMs, post-training quantization (PTQ) can significantly reduce memory footprint and computational overhead. Model quantization is a rapidly evolving research field. Though many papers have reported breakthrough performance, they may not conduct experiments on the same ground since one quantization method usually contains multiple components. In addition, analyzing the theoretical connections among existing methods is crucial for in-depth understanding. To bridge these gaps, we conduct an extensive review of state-of-the-art methods and perform comprehensive evaluations on the same ground to ensure fair comparisons. To our knowledge, this fair and extensive investigation remains critically important yet underexplored. To better understand the theoretical connections, we decouple the published quantization methods into two steps: pre-quantization transformation and quantization error mitigation. We define the former as a preprocessing step applied before quantization to reduce the impact of outliers, making the data distribution flatter and more suitable for quantization. Quantization error mitigation involves techniques that offset the errors introduced during quantization, thereby enhancing model performance. We evaluate and analyze the impact of different components of quantization methods. Additionally, we analyze and evaluate the latest MXFP4 data format and its performance. Our experimental results demonstrate that optimized rotation and scaling yield the best performance for pre-quantization transformation, and combining low-rank compensation with GPTQ occasionally outperforms using GPTQ alone for quantization error mitigation. Furthermore, we explore the potential of the latest MXFP4 quantization and reveal that the optimal pre-quantization transformation strategy for INT4 does not generalize well to MXFP4, inspiring further investigation.