Rethinking AI Quantization: The Missing Piece in Model Efficiency

Authors: (1) Wanyun Cui, Shanghai University of Finance and Economics, with equal contribution; (2) Qianle Wang, Shanghai University of Finance and Economics, with equal contribution. Table of Links Abstract and 1 Introduction 2 Related Work 3 Quantifying the Impact of Parameters on Model Performance & 4. Unified Mixed-Precision Training 5 Prevalence of Parameter Heterogeneity in LLMs 6 Quantization Experiments and 6.1 Implementation Details 6.2 Effect of Base LLM Quantization 6.3 Effect of Chat LLM Quantization 6.4 Comparison of Parameter Selection Criteria, Conclusion, & References 2. Related Work Quantization Strategies for LLMs Various quantization strategies have been proposed in the literature to reduce the precision of weights and activations while maintaining acceptable accuracy. These strategies can be broadly categorized into post-training quantization and quantization-aware training [14]. Post-training quantization methods, such as OBD, OBS, and GPTQ, directly quantize the pre-trained model without fine-tuning [15, 10, 8]. On the other hand, quantization-aware training methods, such as LLM-QAT [18], incorporate quantization operations into the training process to jointly optimize the quantized model. Some works also explore mixed-precision quantization [13] and adaptive quantization bins [7] to achieve a better trade-off between accuracy and efficiency. Outliers in Language Model Quantization The idea of modeling parameter outliers in LLM quantization is not new. The exploration of outliers primarily includes the perspectives of magnitude [18, 7] and activations [4, 6]. For example, from the magnitude perspective, QLoRA assumes that parameters follow a Gaussian distribution [7] and designs information-theoretically optimal quantized bins based on this assumption. [18] keeps outlier parameters in 16-bit precision. From the activation perspective, [17] migrates the outlier amplifier to subsequent modules through an equivalent transformation. Additionally, SqueezeLLM also measures outliers from the perspective of parameter impact [13]. To the best of our knowledge, our work is the first to systematically reveal the outliers (heterogeneity) of parameter impact across different models, and we show a more pronounced imbalance in parameter impacts compared to magnitudes (§ 6.4). Furthermore, we propose a method to unify outlier (cherry) parameter optimization and normal parameter optimization, addressing the optimization challenges of heterogeneous parameters. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Wanyun Cui, Shanghai University of Finance and Economics, with equal contribution; (2) Qianle Wang, Shanghai University of Finance and Economics, with equal contribution. Authors: Authors: (1) Wanyun Cui, Shanghai University of Finance and Economics, with equal contribution; (2) Qianle Wang, Shanghai University of Finance and Economics, with equal contribution. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Quantifying the Impact of Parameters on Model Performance & 4. Unified Mixed-Precision Training 3 Quantifying the Impact of Parameters on Model Performance & 4. Unified Mixed-Precision Training 5 Prevalence of Parameter Heterogeneity in LLMs 5 Prevalence of Parameter Heterogeneity in LLMs 6 Quantization Experiments and 6.1 Implementation Details 6 Quantization Experiments and 6.1 Implementation Details 6.2 Effect of Base LLM Quantization 6.2 Effect of Base LLM Quantization 6.3 Effect of Chat LLM Quantization 6.3 Effect of Chat LLM Quantization 6.4 Comparison of Parameter Selection Criteria, Conclusion, & References 6.4 Comparison of Parameter Selection Criteria, Conclusion, & References 2. Related Work Quantization Strategies for LLMs Various quantization strategies have been proposed in the literature to reduce the precision of weights and activations while maintaining acceptable accuracy. These strategies can be broadly categorized into post-training quantization and quantization-aware training [14]. Post-training quantization methods, such as OBD, OBS, and GPTQ, directly quantize the pre-trained model without fine-tuning [15, 10, 8]. On the other hand, quantization-aware training methods, such as LLM-QAT [18], incorporate quantization operations into the training process to jointly optimize the quantized model. Some works also explore mixed-precision quantization [13] and adaptive quantization bins [7] to achieve a better trade-off between accuracy and efficiency. Quantization Strategies for LLMs Outliers in Language Model Quantization The idea of modeling parameter outliers in LLM quantization is not new. The exploration of outliers primarily includes the perspectives of magnitude [18, 7] and activations [4, 6]. For example, from the magnitude perspective, QLoRA assumes that parameters follow a Gaussian distribution [7] and designs information-theoretically optimal quantized bins based on this assumption. [18] keeps outlier parameters in 16-bit precision. From the activation perspective, [17] migrates the outlier amplifier to subsequent modules through an equivalent transformation. Additionally, SqueezeLLM also measures outliers from the perspective of parameter impact [13]. To the best of our knowledge, our work is the first to systematically reveal the outliers (heterogeneity) of parameter impact across different models, and we show a more pronounced imbalance in parameter impacts compared to magnitudes (§ 6.4). Furthermore, we propose a method to unify outlier (cherry) parameter optimization and normal parameter optimization, addressing the optimization challenges of heterogeneous parameters. Outliers in Language Model Quantization This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv