Key-value (KV) caching is an important technique to accelerate the inference of large language models (LLMs), but incurs significant memory overhead.To compress the size of KV cache, existing methods often compromise precision or require extra data for calibration, limiting their practicality in LLM deployment.In this paper, we introduce DecoQuant, a novel data-free low-bit quantization technique based on tensor decomposition methods, to effectively compress KV cache.Our core idea is to adjust the outlier distribution of the original matrix by performing tensor decomposition, so that the quantization difficulties are migrated from the matrix to decomposed local tensors.Specially, we find that outliers mainly concentrate on small local tensors, while large tensors tend to have a narrower value range.Based on this finding, we propose to apply low-bit quantization to the large tensor, while maintaining high-precision representation for the small tensor.Furthermore, we utilize the proposed quantization method to compress the KV cache of LLMs to accelerate the inference and develop an efficient dequantization kernel tailored specifically for DecoQuant.Through extensive experiments, DecoQuant demonstrates remarkable efficiency gains, showcasing up to a ∼75% reduction in memory footprint while maintaining comparable generation quality.

Lightweight fine-tuning is widely used as an important technique for efficiently adapting pre-trained language models (PLM) to downstream tasks. Despite the reduction in trainable parameters, existing lightweight fine-tuning methods are found to be effective in low-resource settings but often fail in high-resource settings, leading to unreliable outcomes. This limitation can be attributed to inflexible strategies: they identify the parameters of the model to be trained before fine-tuning and remain unchanged without taking into account the inherent variance of generalization ability in model components (i.e., feed-forward, attention layers) and potential changes during the fine-tuning process. In this paper, we introduce a simple but effective calibration for lightweight fine-tuning PLMs based on the matrix’s stable rank according to both model components and the training process. We proposed both theoretical analyses and experimental verification for the proposed calibration strategy. Considering efficiency, we further propose time-aware and structure-aware strategies to determine the most crucial time to commence the fine-tuning procedure and selectively apply parameter matrices for lightweight fine-tuning, respectively. Extensive experiments demonstrate the superiority of our proposed fine-tuning approach (average improvement 3.1 for GLUE score compared to lightweight fine-tuning method).

Despite the superior performance, Large Language Models (LLMs) require significant computational resources for deployment and use. To overcome this issue, quantization methods have been widely applied to reduce the memory footprint of LLMs as well as increase the inference rate. However, a major challenge is that low-bit quantization methods often lead to performance degradation. It is important to understand how quantization impacts the capacity of LLMs. Different from previous studies focused on overall performance, this work aims to investigate the impact of quantization on emergent abilities, which are important characteristics that distinguish LLMs from small language models. Specifically, we examine the abilities of in-context learning, chain-of-thought reasoning, and instruction-following in quantized LLMs. Our empirical experiments show that these emergent abilities still exist in 4-bit quantization models, while 2-bit models encounter severe performance degradation on the test of these abilities. To improve the performance of low-bit models, we conduct two special experiments: (1) fine-gained impact analysis that studies which components (or substructures) are more sensitive to quantization, and (2) performance compensation through model fine-tuning. Our work derives a series of important findings to understand the impact of quantization on emergent abilities and sheds light on the possibilities of extremely low-bit quantization for LLMs.

Increased training parameters have enabled large pre-trained models to excel in various downstream tasks. Nevertheless, the extensive computational requirements associated with these models hinder their widespread adoption within the community. We focus on Knowledge Distillation (KD), where a compact student model is trained to mimic a larger teacher model, facilitating the transfer of knowledge of large models. In contrast to much of the previous work, we scale up the parameters of the student model during training, to benefit from over-parameterization without increasing the inference latency. In particular, we propose a tensor decomposition strategy that effectively over-parameterizes the relatively small student model through an efficient and nearly lossless decomposition of its parameter matrices into higher-dimensional tensors. To ensure efficiency, we further introduce a tensor constraint loss to align the high-dimensional tensors between the student and teacher models. Comprehensive experiments validate the significant performance enhancement by our approach in various KD tasks, covering computer vision and natural language processing areas. Our code is available at https://github.com/intell-sci-comput/OPDF.

By scaling the model size, large pre-trained language models (PLMs) have shown remarkable performance in various natural language processing tasks, mostly outperforming small PLMs by a large margin.However, due to the high computational cost, the huge number of parameters also restricts the applicability of large PLMs in real-world systems.In this paper, we focus on scaling up the parameters of PLMs only during fine-tuning, to benefit from the over-parameterization, while without increasing the inference latency.Given a relatively small PLM, we over-parameterize it by employing a matrix product operator, an efficient and almost lossless decomposition method to factorize its contained parameter matrices into a set of higher-dimensional tensors.Considering the efficiency, we further propose both static and dynamic strategies to select the most important parameter matrices for over-parameterization.Extensive experiments have demonstrated that our approach can significantly boost the finetuning performance of small PLMs and even help small PLMs outperform 3× parameterized larger ones.

Recently, Mixture-of-Experts (short as MoE) architecture has achieved remarkable success in increasing the model capacity of large-scale language models. However, MoE requires incorporating significantly more parameters than the base model being extended. In this paper, we propose building a parameter-efficient MoE architecture by sharing information across experts. We adopt matrix product operator (MPO, a tensor decomposition from quantum many-body physics) to reconstruct the parameter matrix in the expert layer and increase model capacity for pre-trained language models by sharing parameters of the central tensor (containing the core information) among different experts while enabling the specificity through the auxiliary tensors (complementing the central tensor) of different experts. To address the unbalanced optimization issue, we further design the gradient mask strategy for the MPO-based MoE architecture. Extensive experiments based on T5 and GPT-2 show improved performance and efficiency of the pre-trained language model (27.2x reduction in total parameters for the superior model performance, compared with the Switch Transformers). Our code is publicly available at https://github.com/RUCAIBox/MPO/MPOE.

Peiyu Liu, Ze-Feng Gao, Wayne Xin Zhao, Zhi-Yuan Xie, Zhong-Yi Lu, Ji-Rong Wen. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

The deep neural network (DNN) based speech enhancement approaches have achieved promising performance. However, the number of parameters involved in these methods is usually enormous for the real applications of speech enhancement on the device with the limited resources. This seriously restricts the applications. To deal with this issue, model compression techniques are being widely studied. In this paper, we propose a model compression method based on matrix product operators (MPO) to substantially reduce the number of parameters in DNN models for speech enhancement. In this method, the weight matrices in the linear transformations of neural network model are replaced by the MPO decomposition format before training. In experiment, this process is applied to the causal neural network models, such as the feedforward multilayer perceptron (MLP) and long short-term memory (LSTM) models. Both MLP and LSTM models with/without compression are then utilized to estimate the ideal ratio mask for monaural speech enhancement. The experimental results show that our proposed MPO-based method outperforms the widely-used pruning method for speech enhancement under various compression rates, and further improvement can be achieved with respect to low compression rates. Our proposal provides an effective model compression method for speech enhancement, especially in cloud-free application.