A critical approach for efficiently deploying computationally demanding large language models (LLMs) is Key-Value (KV) caching. The KV cache stores key-value states of previously generated tokens, significantly reducing the need for repetitive computations and thereby lowering latency in autoregressive generation. However, the size of the KV cache grows linearly with sequence length, posing challenges for applications requiring long context input and extensive sequence generation. In this paper, we present a simple yet effective approach, called MiniCache, to compress the KV cache across layers from a novel depth perspective, significantly reducing the memory footprint for LLM inference. Our approach is based on the observation that KV cache states exhibit high similarity between the adjacent layers in the middle-to-deep portion of LLMs. To facilitate merging, we propose disentangling the states into the magnitude and direction components, interpolating the directions of the state vectors while preserving their lengths unchanged. Furthermore, we introduce a token retention strategy to keep highly distinct state pairs unmerged, thus preserving the information with minimal additional storage overhead. Our MiniCache is training-free and general, complementing existing KV cache compression strategies, such as quantization and sparsity. We conduct a comprehensive evaluation of MiniCache utilizing various models including LLaMA-2, LLaMA-3, Phi-3, Mistral, and Mixtral across multiple benchmarks, demonstrating its exceptional performance in achieving superior compression ratios and high throughput. On the ShareGPT dataset, LLaMA-2-7B with 4-bit MiniCache achieves a remarkable compression ratio of up to 5.02x, enhances inference throughput by approximately 5x, and reduces the memory footprint by 41% compared to the FP16 full cache baseline, all while maintaining near-lossless performance.
KV cache stores key and value states from previous tokens to avoid re-computation, yet it demands substantial storage space, especially for long sequences. Adaptive KV cache compression seeks to discern the saliency of tokens, preserving vital information while aggressively compressing those of less importance. However, previous methods of this approach exhibit significant performance degradation at high compression ratios due to inaccuracies in identifying salient tokens. In this paper, we present ZipCache, an accurate and efficient KV cache quantization method for LLMs. First, we construct a strong baseline for quantizing KV cache. Through the proposed channel-separable tokenwise quantization scheme, the memory overhead of quantization parameters are substantially reduced compared to fine-grained groupwise quantization. To enhance the compression ratio, we propose normalized attention score as an effective metric for identifying salient tokens by considering the lower triangle characteristics of the attention matrix. Moreover, we develop an efficient approximation method that decouples the saliency metric from full attention scores, enabling compatibility with fast attention implementations like FlashAttention. Extensive experiments demonstrate that ZipCache achieves superior compression ratios, fast generation speed and minimal performance losses compared with previous KV cache compression methods. For instance, when evaluating Mistral-7B model on GSM8k dataset, ZipCache is capable of compressing the KV cache by $4.98\times$, with only a $0.38\%$ drop in accuracy. In terms of efficiency, ZipCache also showcases a $37.3\%$ reduction in prefill-phase latency, a $56.9\%$ reduction in decoding-phase latency, and a $19.8\%$ reduction in GPU memory usage when evaluating LLaMA3-8B model with a input length of $4096$.
Quantization is widely used to increase deep neural networks' (DNN) memory, computation, and power efficiency. Various techniques, such as post-training quantization and quantization-aware training, have been proposed to improve quantization quality. We introduce a novel approach for DNN quantization that uses a redundant representation of DNN's output. We represent the target quantity as a point on a 2D parametric curve. The DNN model is modified to predict 2D points that are mapped back to the target quantity at a post-processing stage. We demonstrate that this mapping can reduce quantization error. For the low-order parametric Hilbert curve, Depth-From-Stereo task, and two models represented by U-Net architecture and vision transformer, we achieved a quantization error reduction by about 5 times for the INT8 model at both CPU and DSP delegates. This gain comes with a minimal inference time increase (less than 7%). Our approach can be applied to other tasks, including segmentation, object detection, and key-points prediction.
Recent developments in large-scale pre-trained text-to-image diffusion models have significantly improved the generation of high-fidelity images, particularly with the emergence of diffusion models based on transformer architecture (DiTs). Among these diffusion models, diffusion transformers have demonstrated superior image generation capabilities, boosting lower FID scores and higher scalability. However, deploying large-scale DiT models can be expensive due to their extensive parameter numbers. Although existing research has explored efficient deployment techniques for diffusion models such as model quantization, there is still little work concerning DiT-based models. To tackle this research gap, in this paper, we propose TerDiT, a quantization-aware training (QAT) and efficient deployment scheme for ternary diffusion models with transformers. We focus on the ternarization of DiT networks and scale model sizes from 600M to 4.2B. Our work contributes to the exploration of efficient deployment strategies for large-scale DiT models, demonstrating the feasibility of training extremely low-bit diffusion transformer models from scratch while maintaining competitive image generation capacities compared to full-precision models. Code will be available at https://github.com/Lucky-Lance/TerDiT.
Predicting future values in multivariate time series is vital across various domains. This work explores the use of large language models (LLMs) for this task. However, LLMs typically handle one-dimensional data. We introduce MultiCast, a zero-shot LLM-based approach for multivariate time series forecasting. It allows LLMs to receive multivariate time series as input, through three novel token multiplexing solutions that effectively reduce dimensionality while preserving key repetitive patterns. Additionally, a quantization scheme helps LLMs to better learn these patterns, while significantly reducing token use for practical applications. We showcase the performance of our approach in terms of RMSE and execution time against state-of-the-art approaches on three real-world datasets.
We introduce Integer Scale, a novel post-training quantization scheme for large language models that effectively resolves the inference bottleneck in current fine-grained quantization approaches while maintaining similar accuracies. Integer Scale is a free lunch as it requires no extra calibration or fine-tuning which will otherwise incur additional costs. It can be used plug-and-play for most fine-grained quantization methods. Its integration results in at most 1.85x end-to-end speed boost over the original counterpart with comparable accuracy. Additionally, due to the orchestration of the proposed Integer Scale and fine-grained quantization, we resolved the quantization difficulty for Mixtral-8x7B and LLaMA-3 models with negligible performance degradation, and it comes with an end-to-end speed boost of 2.13x, and 2.31x compared with their FP16 versions respectively.
Unrolled deep neural networks have attracted significant attention for their success in various practical applications. In this paper, we explore an application of deep unrolling in the direction of arrival (DoA) estimation problem when coarse quantization is applied to the measurements. We present a compressed sensing formulation for DoA estimation from one-bit data in which estimating target DoAs requires recovering a sparse signal from a limited number of severely quantized linear measurements. In particular, we exploit covariance recovery from one-bit dither samples. To recover the covariance of transmitted signal, the learned iterative shrinkage and thresholding algorithm (LISTA) is employed fed by one-bit data. We demonstrate that the upper bound of estimation performance is governed by the recovery error of the transmitted signal covariance matrix. Through numerical experiments, we demonstrate the proposed LISTA-based algorithm's capability in estimating target locations. The code employed in this study is available online.
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 \textbf{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 $\sim$75\% reduction in memory footprint while maintaining comparable generation quality.
Searching for approximate nearest neighbors (ANN) in the high-dimensional Euclidean space is a pivotal problem. Recently, with the help of fast SIMD-based implementations, Product Quantization (PQ) and its variants can often efficiently and accurately estimate the distances between the vectors and have achieved great success in the in-memory ANN search. Despite their empirical success, we note that these methods do not have a theoretical error bound and are observed to fail disastrously on some real-world datasets. Motivated by this, we propose a new randomized quantization method named RaBitQ, which quantizes $D$-dimensional vectors into $D$-bit strings. RaBitQ guarantees a sharp theoretical error bound and provides good empirical accuracy at the same time. In addition, we introduce efficient implementations of RaBitQ, supporting to estimate the distances with bitwise operations or SIMD-based operations. Extensive experiments on real-world datasets confirm that (1) our method outperforms PQ and its variants in terms of accuracy-efficiency trade-off by a clear margin and (2) its empirical performance is well-aligned with our theoretical analysis.
The advancement of sophisticated artificial intelligence (AI) algorithms has led to a notable increase in energy usage and carbon dioxide emissions, intensifying concerns about climate change. This growing problem has brought the environmental sustainability of AI technologies to the forefront, especially as they expand across various sectors. In response to these challenges, there is an urgent need for the development of sustainable AI solutions. These solutions must focus on energy-efficient embedded systems that are capable of handling diverse data types even in environments with limited resources, thereby ensuring both technological progress and environmental responsibility. Integrating complementary multimodal data into tiny machine learning models for edge devices is challenging due to increased complexity, latency, and power consumption. This work introduces TinyM$^2$Net-V3, a system that processes different modalities of complementary data, designs deep neural network (DNN) models, and employs model compression techniques including knowledge distillation and low bit-width quantization with memory-aware considerations to fit models within lower memory hierarchy levels, reducing latency and enhancing energy efficiency on resource-constrained devices. We evaluated TinyM$^2$Net-V3 in two multimodal case studies: COVID-19 detection using cough, speech, and breathing audios, and pose classification from depth and thermal images. With tiny inference models (6 KB and 58 KB), we achieved 92.95% and 90.7% accuracies, respectively. Our tiny machine learning models, deployed on resource limited hardware, demonstrated low latencies within milliseconds and very high power efficiency.
Model quantization is widely used to compress and accelerate deep neural networks. However, recent studies have revealed the feasibility of weaponizing model quantization via implanting quantization-conditioned backdoors (QCBs). These special backdoors stay dormant on released full-precision models but will come into effect after standard quantization. Due to the peculiarity of QCBs, existing defenses have minor effects on reducing their threats or are even infeasible. In this paper, we conduct the first in-depth analysis of QCBs. We reveal that the activation of existing QCBs primarily stems from the nearest rounding operation and is closely related to the norms of neuron-wise truncation errors (i.e., the difference between the continuous full-precision weights and its quantized version). Motivated by these insights, we propose Error-guided Flipped Rounding with Activation Preservation (EFRAP), an effective and practical defense against QCBs. Specifically, EFRAP learns a non-nearest rounding strategy with neuron-wise error norm and layer-wise activation preservation guidance, flipping the rounding strategies of neurons crucial for backdoor effects but with minimal impact on clean accuracy. Extensive evaluations on benchmark datasets demonstrate that our EFRAP can defeat state-of-the-art QCB attacks under various settings. Code is available at https://github.com/AntigoneRandy/QuantBackdoor_EFRAP.
This research introduces an innovative approach to explore the cognitive and biologically inspired underpinnings of feature vector splitting for analyzing the significance of different attributes in e-security biometric signature recognition applications. Departing from traditional methods of concatenating features into an extended set, we employ multiple splitting strategies, aligning with cognitive principles, to preserve control over the relative importance of each feature subset. Our methodology is applied to three diverse databases (MCYT100, MCYT300,and SVC) using two classifiers (vector quantization and dynamic time warping with one and five training samples). Experimentation demonstrates that the fusion of pressure data with spatial coordinates (x and y) consistently enhances performance. However, the inclusion of pen-tip angles in the same feature set yields mixed results, with performance improvements observed in select cases. This work delves into the cognitive aspects of feature fusion,shedding light on the cognitive relevance of feature vector splitting in e-security biometric applications.
Image registration is a classical problem in machine vision which seeks methods to align discrete images of the same scene to subpixel accuracy in general situations. As with all estimation problems, the underlying difficulty is the partial information available about the ground truth. We consider a basic and idealized one-dimensional image registration problem motivated by questions about measurement and about quantization, and we demonstrate that the extent to which subinterval/subpixel inferences can be made in this setting depends on a type of complexity associated with the function of interest, the relationship between the function and the pixel size, and the number of distinct sampling count observations available.
This paper provides a comprehensive review of the future of cybersecurity through Generative AI and Large Language Models (LLMs). We explore LLM applications across various domains, including hardware design security, intrusion detection, software engineering, design verification, cyber threat intelligence, malware detection, and phishing detection. We present an overview of LLM evolution and its current state, focusing on advancements in models such as GPT-4, GPT-3.5, Mixtral-8x7B, BERT, Falcon2, and LLaMA. Our analysis extends to LLM vulnerabilities, such as prompt injection, insecure output handling, data poisoning, DDoS attacks, and adversarial instructions. We delve into mitigation strategies to protect these models, providing a comprehensive look at potential attack scenarios and prevention techniques. Furthermore, we evaluate the performance of 42 LLM models in cybersecurity knowledge and hardware security, highlighting their strengths and weaknesses. We thoroughly evaluate cybersecurity datasets for LLM training and testing, covering the lifecycle from data creation to usage and identifying gaps for future research. In addition, we review new strategies for leveraging LLMs, including techniques like Half-Quadratic Quantization (HQQ), Reinforcement Learning with Human Feedback (RLHF), Direct Preference Optimization (DPO), Quantized Low-Rank Adapters (QLoRA), and Retrieval-Augmented Generation (RAG). These insights aim to enhance real-time cybersecurity defenses and improve the sophistication of LLM applications in threat detection and response. Our paper provides a foundational understanding and strategic direction for integrating LLMs into future cybersecurity frameworks, emphasizing innovation and robust model deployment to safeguard against evolving cyber threats.
We propose a novel algorithm for distributed stochastic gradient descent (SGD) with compressed gradient communication in the parameter-server framework. Our gradient compression technique, named flattened one-bit stochastic gradient descent (FO-SGD), relies on two simple algorithmic ideas: (i) a one-bit quantization procedure leveraging the technique of dithering, and (ii) a randomized fast Walsh-Hadamard transform to flatten the stochastic gradient before quantization. As a result, the approximation of the true gradient in this scheme is biased, but it prevents commonly encountered algorithmic problems, such as exploding variance in the one-bit compression regime, deterioration of performance in the case of sparse gradients, and restrictive assumptions on the distribution of the stochastic gradients. In fact, we show SGD-like convergence guarantees under mild conditions. The compression technique can be used in both directions of worker-server communication, therefore admitting distributed optimization with full communication compression.
By harnessing the capabilities of large language models (LLMs), recent large multimodal models (LMMs) have shown remarkable versatility in open-world multimodal understanding. Nevertheless, they are usually parameter-heavy and computation-intensive, thus hindering their applicability in resource-constrained scenarios. To this end, several lightweight LMMs have been proposed successively to maximize the capabilities under constrained scale (e.g., 3B). Despite the encouraging results achieved by these methods, most of them only focus on one or two aspects of the design space, and the key design choices that influence model capability have not yet been thoroughly investigated. In this paper, we conduct a systematic study for lightweight LMMs from the aspects of model architecture, training strategy, and training data. Based on our findings, we obtain Imp -- a family of highly capable LMMs at the 2B-4B scales. Notably, our Imp-3B model steadily outperforms all the existing lightweight LMMs of similar size, and even surpasses the state-of-the-art LMMs at the 13B scale. With low-bit quantization and resolution reduction techniques, our Imp model can be deployed on a Qualcomm Snapdragon 8Gen3 mobile chip with a high inference speed of about 13 tokens/s.
Similar to natural language models, pre-trained genome language models are proposed to capture the underlying intricacies within genomes with unsupervised sequence modeling. They have become essential tools for researchers and practitioners in biology. However, the hand-crafted tokenization policies used in these models may not encode the most discriminative patterns from the limited vocabulary of genomic data. In this paper, we introduce VQDNA, a general-purpose framework that renovates genome tokenization from the perspective of genome vocabulary learning. By leveraging vector-quantized codebooks as learnable vocabulary, VQDNA can adaptively tokenize genomes into pattern-aware embeddings in an end-to-end manner. To further push its limits, we propose Hierarchical Residual Quantization (HRQ), where varying scales of codebooks are designed in a hierarchy to enrich the genome vocabulary in a coarse-to-fine manner. Extensive experiments on 32 genome datasets demonstrate VQDNA's superiority and favorable parameter efficiency compared to existing genome language models. Notably, empirical analysis of SARS-CoV-2 mutations reveals the fine-grained pattern awareness and biological significance of learned HRQ vocabulary, highlighting its untapped potential for broader applications in genomics.
From the perspective of joint source-channel coding (JSCC), there has been significant research on utilizing semantic communication, which inherently possesses analog characteristics, within digital device environments. However, a single-model approach that operates modulation-agnostically across various digital modulation orders has not yet been established. This article presents the first attempt at such an approach by proposing a universal joint source-channel coding (uJSCC) system that utilizes a single-model encoder-decoder pair and trained vector quantization (VQ) codebooks. To support various modulation orders within a single model, the operation of every neural network (NN)-based module in the uJSCC system requires the selection of modulation orders according to signal-to-noise ratio (SNR) boundaries. To address the challenge of unequal output statistics from shared parameters across NN layers, we integrate multiple batch normalization (BN) layers, selected based on modulation order, after each NN layer. This integration occurs with minimal impact on the overall model size. Through a comprehensive series of experiments, we validate that the modulation-agnostic semantic communication framework demonstrates superiority over existing digital semantic communication approaches in terms of model complexity, communication efficiency, and task effectiveness.
Decoding remote sensing images to achieve high perceptual quality, particularly at low bitrates, remains a significant challenge. To address this problem, we propose the invertible neural network-based remote sensing image compression (INN-RSIC) method. Specifically, we capture compression distortion from an existing image compression algorithm and encode it as a set of Gaussian-distributed latent variables via INN. This ensures that the compression distortion in the decoded image becomes independent of the ground truth. Therefore, by leveraging the inverse mapping of INN, we can input the decoded image along with a set of randomly resampled Gaussian distributed variables into the inverse network, effectively generating enhanced images with better perception quality. To effectively learn compression distortion, channel expansion, Haar transformation, and invertible blocks are employed to construct the INN. Additionally, we introduce a quantization module (QM) to mitigate the impact of format conversion, thus enhancing the framework's generalization and improving the perceptual quality of enhanced images. Extensive experiments demonstrate that our INN-RSIC significantly outperforms the existing state-of-the-art traditional and deep learning-based image compression methods in terms of perception quality.
Federated Learning (FL) is an emerging paradigm that holds great promise for privacy-preserving machine learning using distributed data. To enhance privacy, FL can be combined with Differential Privacy (DP), which involves adding Gaussian noise to the model weights. However, FL faces a significant challenge in terms of large communication overhead when transmitting these model weights. To address this issue, quantization is commonly employed. Nevertheless, the presence of quantized Gaussian noise introduces complexities in understanding privacy protection. This research paper investigates the impact of quantization on privacy in FL systems. We examine the privacy guarantees of quantized Gaussian mechanisms using R\'enyi Differential Privacy (RDP). By deriving the privacy budget of quantized Gaussian mechanisms, we demonstrate that lower quantization bit levels provide improved privacy protection. To validate our theoretical findings, we employ Membership Inference Attacks (MIA), which gauge the accuracy of privacy leakage. The numerical results align with our theoretical analysis, confirming that quantization can indeed enhance privacy protection. This study not only enhances our understanding of the correlation between privacy and communication in FL but also underscores the advantages of quantization in preserving privacy.