BitDistiller: Unleashing the Potential of Sub-4-Bit LLMs via Self-Distillation

The upscaling of Large Language Models (LLMs) has yielded impressive advances in natural language processing, yet it also poses significant deployment challenges. Weight quantization has emerged as a widely embraced solution to reduce memory and computational demands. This paper introduces BitDistiller, a framework that synergizes Quantization-Aware Training (QAT) with Knowledge Distillation (KD) to boost the performance of LLMs at ultra-low precisions (sub-4-bit). Specifically, BitDistiller first incorporates a tailored asymmetric quantization and clipping technique to maximally preserve the fidelity of quantized weights, and then proposes a novel Confidence-Aware Kullback-Leibler Divergence (CAKLD) objective, which is employed in a self-distillation manner to enable faster convergence and superior model performance. Empirical evaluations demonstrate that BitDistiller significantly surpasses existing methods in both 3-bit and 2-bit configurations on general language understanding and complex reasoning benchmarks. Notably, BitDistiller is shown to be more cost-effective, demanding fewer data and training resources. The code is available at https://github.com/DD-DuDa/BitDistiller.

CharacterGLM: Customizing Chinese Conversational AI Characters with Large Language Models

In this paper, we present CharacterGLM, a series of models built upon ChatGLM, with model sizes ranging from 6B to 66B parameters. Our CharacterGLM is designed for generating Character-based Dialogues (CharacterDial), which aims to equip a conversational AI system with character customization for satisfying people's inherent social desires and emotional needs. On top of CharacterGLM, we can customize various AI characters or social agents by configuring their attributes (identities, interests, viewpoints, experiences, achievements, social relationships, etc.) and behaviors (linguistic features, emotional expressions, interaction patterns, etc.). Our model outperforms most mainstream close-source large langauge models, including the GPT series, especially in terms of consistency, human-likeness, and engagement according to manual evaluations. We will release our 6B version of CharacterGLM and a subset of training data to facilitate further research development in the direction of character-based dialogue generation.

AFPQ: Asymmetric Floating Point Quantization for LLMs

Large language models (LLMs) show great performance in various tasks, but face deployment challenges from limited memory capacity and bandwidth. Low-bit weight quantization can save memory and accelerate inference. Although floating-point (FP) formats show good performance in LLM quantization, they tend to perform poorly with small group sizes or sub-4 bits. We find the reason is that the absence of asymmetry in previous FP quantization makes it unsuitable for handling asymmetric value distribution of LLM weight tensors. In this work, we propose asymmetric FP quantization (AFPQ), which sets separate scales for positive and negative values. Our method leads to large accuracy improvements and can be easily plugged into other quantization methods, including GPTQ and AWQ, for better performance. Besides, no additional storage is needed compared with asymmetric integer (INT) quantization. The code is available at https://github.com/zhangsichengsjtu/AFPQ.

A Transformer-Based Model With Self-Distillation for Multimodal Emotion Recognition in Conversations

Emotion recognition in conversations (ERC), the task of recognizing the emotion of each utterance in a conversation, is crucial for building empathetic machines. Existing studies focus mainly on capturing context- and speaker-sensitive dependencies on the textual modality but ignore the significance of multimodal information. Different from emotion recognition in textual conversations, capturing intra- and inter-modal interactions between utterances, learning weights between different modalities, and enhancing modal representations play important roles in multimodal ERC. In this paper, we propose a transformer-based model with self-distillation (SDT) for the task. The transformer-based model captures intra- and inter-modal interactions by utilizing intra- and inter-modal transformers, and learns weights between modalities dynamically by designing a hierarchical gated fusion strategy. Furthermore, to learn more expressive modal representations, we treat soft labels of the proposed model as extra training supervision. Specifically, we introduce self-distillation to transfer knowledge of hard and soft labels from the proposed model to each modality. Experiments on IEMOCAP and MELD datasets demonstrate that SDT outperforms previous state-of-the-art baselines.

TKwinFormer: Top k Window Attention in Vision Transformers for Feature Matching

Local feature matching remains a challenging task, primarily due to difficulties in matching sparse keypoints and low-texture regions. The key to solving this problem lies in effectively and accurately integrating global and local information. To achieve this goal, we introduce an innovative local feature matching method called TKwinFormer. Our approach employs a multi-stage matching strategy to optimize the efficiency of information interaction. Furthermore, we propose a novel attention mechanism called Top K Window Attention, which facilitates global information interaction through window tokens prior to patch-level matching, resulting in improved matching accuracy. Additionally, we design an attention block to enhance attention between channels. Experimental results demonstrate that TKwinFormer outperforms state-of-the-art methods on various benchmarks. Code is available at: https://github.com/LiaoYun0x0/TKwinFormer.

Adam Accumulation to Reduce Memory Footprints of both Activations and Gradients for Large-scale DNN Training

Running out of GPU memory has become a main bottleneck for large-scale DNN training. How to reduce the memory footprint during training has received intensive research attention. We find that previous gradient accumulation reduces activation memory but fails to be compatible with gradient memory reduction due to a contradiction between preserving gradients and releasing gradients. To address this issue, we propose a novel optimizer accumulation method for Adam, named Adam Accumulation (AdamA), which enables reducing both activation and gradient memory. Specifically, AdamA directly integrates gradients into optimizer states and accumulates optimizer states over micro-batches, so that gradients can be released immediately after use. We mathematically and experimentally demonstrate AdamA yields the same convergence properties as Adam. Evaluated on transformer-based models, AdamA achieves up to 23% memory reduction compared to gradient accumulation with less than 2% degradation in training throughput. Notably, AdamA can work together with memory reduction methods for optimizer states to fit 1.26x~3.14x larger models over PyTorch and DeepSpeed baseline on GPUs with different memory capacities.

Integer or Floating Point? New Outlooks for Low-Bit Quantization on Large Language Models

Efficient deployment of large language models (LLMs) necessitates low-bit quantization to minimize model size and inference cost. While low-bit integer formats (e.g., INT8/INT4) have been the conventional choice, emerging low-bit floating-point formats (e.g., FP8/FP4) offer a compelling alternative and are gaining support from cutting-edge hardware, such as NVIDIA's H100 GPU. However, the superiority of low-bit INT versus FP formats for quantization on LLMs remains unclear. In this study, we conduct a comparative analysis of INT and FP quantization with the same bit-width, revealing that the optimal quantization format varies across different layers due to the complexity and diversity of tensor distribution. Consequently, we advocate the Mixture of Formats Quantization (MoFQ), which selects the optimal format on a layer-wise basis. This simple yet effective approach achieves state-of-the-art results in both weight-only (W-only) and weight-activation (WA) post-training quantization scenarios when tested on LLaMA across various tasks. In 4-bit W-only quantization, MoFQ surpasses GPTQ without complex hyperparameter tuning and with an order of magnitude faster quantization speed. While in 8-bit WA quantization, MoFQ significantly outperforms INT/FP-only methods, achieving performance close to the full precision model. Notably, MoFQ incurs no hardware overhead compared to INT/FP-only quantization, as the bit-width remains unchanged.

TC-GAT: Graph Attention Network for Temporal Causality Discovery

The present study explores the intricacies of causal relationship extraction, a vital component in the pursuit of causality knowledge. Causality is frequently intertwined with temporal elements, as the progression from cause to effect is not instantaneous but rather ensconced in a temporal dimension. Thus, the extraction of temporal causality holds paramount significance in the field. In light of this, we propose a method for extracting causality from the text that integrates both temporal and causal relations, with a particular focus on the time aspect. To this end, we first compile a dataset that encompasses temporal relationships. Subsequently, we present a novel model, TC-GAT, which employs a graph attention mechanism to assign weights to the temporal relationships and leverages a causal knowledge graph to determine the adjacency matrix. Additionally, we implement an equilibrium mechanism to regulate the interplay between temporal and causal relations. Our experiments demonstrate that our proposed method significantly surpasses baseline models in the task of causality extraction.

Exploring Semi-supervised Variational Autoencoders for Biomedical Relation Extraction

The biomedical literature provides a rich source of knowledge such as protein-protein interactions (PPIs), drug-drug interactions (DDIs) and chemical-protein interactions (CPIs). Biomedical relation extraction aims to automatically extract biomedical relations from biomedical text for various biomedical research. State-of-the-art methods for biomedical relation extraction are primarily based on supervised machine learning and therefore depend on (sufficient) labeled data. However, creating large sets of training data is prohibitively expensive and labor-intensive, especially so in biomedicine as domain knowledge is required. In contrast, there is a large amount of unlabeled biomedical text available in PubMed. Hence, computational methods capable of employing unlabeled data to reduce the burden of manual annotation are of particular interest in biomedical relation extraction. We present a novel semi-supervised approach based on variational autoencoder (VAE) for biomedical relation extraction. Our model consists of the following three parts, a classifier, an encoder and a decoder. The classifier is implemented using multi-layer convolutional neural networks (CNNs), and the encoder and decoder are implemented using both bidirectional long short-term memory networks (Bi-LSTMs) and CNNs, respectively. The semi-supervised mechanism allows our model to learn features from both the labeled and unlabeled data. We evaluate our method on multiple public PPI, DDI and CPI corpora. Experimental results show that our method effectively exploits the unlabeled data to improve the performance and reduce the dependence on labeled data. To our best knowledge, this is the first semi-supervised VAE-based method for (biomedical) relation extraction. Our results suggest that exploiting such unlabeled data can be greatly beneficial to improved performance in various biomedical relation extraction.