LCQ: Low-Rank Codebook based Quantization for Large Language Models
Wen-Pu Cai, Ming-Yang Li, Wu-Jun Li
TL;DR
This work tackles the challenge of deploying large language models under storage and compute constraints by enhancing weight quantization. It introduces LCQ, a low-rank codebook quantization method where the codebook is $\mathbf{C} = \mathbf{S}^T \mathbf{V} - \mathbf{B}$, allowing ranks greater than one for richer representation, and optimizes $\mathbf{S}, \mathbf{V}, \mathbf{B}$ via gradient-based learning. The framework uses a Transformer-wide output-reconstruction objective, gradient approximations for quantization, and reparameterization to stabilize training, complemented by a double-quantization strategy to reduce storage. Empirical results across OPT, LLaMA, and LLaVA show LCQ outperforms rank-one baselines (AWQ, OmniQuant), especially at 2-bit quantization, with negligible additional storage, supporting practical deployment on resource-constrained devices. The approach advances LLM quantization by balancing accuracy, storage, and compatibility with existing PTQ workflows, enabling more efficient and accessible large-scale models.
Abstract
Large language models~(LLMs) have recently demonstrated promising performance in many tasks. However, the high storage and computational cost of LLMs has become a challenge for deploying LLMs. Weight quantization has been widely used for model compression, which can reduce both storage and computational cost. Most existing weight quantization methods for LLMs use a rank-one codebook for quantization, which results in substantial accuracy loss when the compression ratio is high. In this paper, we propose a novel weight quantization method, called low-rank codebook based quantization~(LCQ), for LLMs. LCQ adopts a low-rank codebook, the rank of which can be larger than one, for quantization. Experiments show that LCQ can achieve better accuracy than existing methods with a negligibly extra storage cost.