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

In this paper, we present CAT-Q, **C**ost-efficient and **A**ccurate **T**ernary **Q**uantization, for compressing and accelerating LLMs. Unlike existing state-of-the-art ternary quantization methods that rely on data-intensive and costly quantization-aware training to mitigate severe performance degradation, CAT-Q is a simple yet effective post-training quantization scheme that is readily applicable to LLMs with diverse architectures and model sizes. It has two key components, learnable modulation (LM) and softened ternarization (ST), which are coupled from an optimization perspective. LM leverages a composition of learnable factors to modulate the distribution of pre-trained high-precision weights and the ternary threshold, making them less sensitive to ternarization. ST further introduces a differentiable transition function to guide the ternarization process toward stable convergence. We show that, for pre-trained LLMs with 1.7B to 8B parameters, CAT-Q can efficiently quantize them into ternary models using only 512 calibration samples, while achieving superior performance than the seminal BitNet 1.58-bit v1 and v2 families (with 1.3B to 7B parameters) trained with 100B tokens, yielding about a 100,000$\times$ reduction in training tokens. Moreover, we show for the first time that CAT-Q can quantize much larger pre-trained LLMs having 14B to 235B parameters into leading ternary models within just 8 to 60 hours on 8 A100-80GB GPUs. Code is available at https://github.com/IntelChina-AI/BitTern.

Large language models (LLMs) are powerful but expensive to store and deploy because they typically contain billions to hundreds of billions of parameters. A promising way to reduce this cost is ternary quantization, which replaces the high-precision values of model parameters with ternary values {1,0,-1}. However, existing ternary quantization methods often require massive retraining data and significant computational resources to avoid serious model accuracy drop. We introduce CAT-Q, a simple and accurate ternary quantization method for efficiently compressing pre-trained LLMs without retraining from scratch. CAT-Q establishes a new framework to guide the conversion of the pre-trained model parameter values into a much more compact ternary form, enabling the compressed model to preserve strong performance using only a small amount of calibration data. In our experiments, CAT-Q compresses LLMs using only 512 calibration samples, while previous top-performing ternary LLM families rely on around 100 billion training tokens. This corresponds to a reduction of around 100,000× in training tokens. More importantly, CAT-Q can easily scale to compress much larger LLMs, ranging from 14B to 235B parameters, within 8 to 60 hours on 8 A100 GPUs. These results demonstrate that highly compact ternary LLMs can be produced with substantially greater efficiency and practicality than ever before.