CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs

Overview

CAT-Q (Cost-efficient and Accurate Ternary Quantization) addresses the challenge of compressing Large Language Models (LLMs) into 1.58-bit ternary weights ({-1, 0, 1}). While existing state-of-the-art ternary methods rely on expensive Quantization-Aware Training (QAT) that requires processing hundreds of billions of tokens, CAT-Q operates as a Post-Training Quantization (PTQ) scheme. This allows it to achieve high-quality compression using only 512 calibration samples, significantly reducing the computational and data requirements for deployment.

Key Components

CAT-Q introduces two primary innovations to stabilize the difficult optimization process of ternary quantization:

• Learnable Modulation (LM): This component uses a set of learnable factors to adjust the distribution of pre-trained weights and the ternary threshold. By transforming the weight distribution before quantization, it suppresses the impact of outliers and reduces the distributional misalignment between high-precision weights and their ternary approximations.
• Softened Ternarization (ST): To overcome the non-differentiable nature of hard ternarization, ST employs a two-stage process. It uses a differentiable transition function that gradually evolves from an identity mapping to a hard ternary output. This smooth transition guides the optimization toward stable convergence during the calibration phase.

Performance and Scalability

Experiments demonstrate that CAT-Q outperforms seminal ternary models like BitNet 1.58-bit v1 and v2 while requiring 100,000 times fewer training tokens. The authors successfully applied the method to a wide range of models, from 1.7B to 235B parameters, including both dense and Mixture-of-Experts (MoE) architectures. The entire quantization process for these large models takes between 8 and 60 hours on 8 A100 GPUs, making it a highly practical solution for industrial deployment.