How to Parameterize Asymmetric Quantization Ranges for Quantization-Aware Training
This work addresses the challenge of optimizing quantization-aware training for machine learning practitioners, particularly in large language models, but it appears incremental as it compares existing parameterizations rather than introducing a new paradigm.
This paper tackled the problem of parameterizing asymmetric quantization ranges for quantization-aware training by comparing three methods (scale and offset, minimum and maximum, beta and gamma) and analyzing their effects on training stability and speed. The result includes proposed best practices to improve quantization-aware training, though no concrete numbers are provided in the abstract.
This paper investigates three different parameterizations of asymmetric uniform quantization for quantization-aware training: (1) scale and offset, (2) minimum and maximum, and (3) beta and gamma. We perform a comprehensive comparative analysis of these parameterizations' influence on quantization-aware training, using both controlled experiments and real-world large language models. Our particular focus is on their changing behavior in response to critical training hyperparameters, bit width and learning rate. Based on our investigation, we propose best practices to stabilize and accelerate quantization-aware training with learnable asymmetric quantization ranges.