Large language models (LLMs) face the challenges in fine-tuning and deployment due to their high memory demands and computational costs. While parameter-efficient fine-tuning (PEFT) methods aim to reduce the memory usage of the optimizer state during fine-tuning, the inherent size of pre-trained LLM weights continues to be a pressing concern. Even though quantization techniques are widely proposed to ease memory demands and accelerate LLM inference, most of these techniques are geared towards the deployment phase. To bridge this gap, this paper presents Parameter-Efficient and Quantization-aware Adaptation (PEQA) - a simple yet effective method that combines the advantages of PEFT with quantized LLMs. By updating solely the quantization scales, PEQA can be directly applied to quantized LLMs, ensuring seamless task transitions. Parallel to existing PEFT methods, PEQA significantly reduces the memory overhead associated with the optimizer state. Furthermore, it leverages the advantages of quantization to substantially reduce model sizes. Even after fine-tuning, the quantization structure of a PEQA-tuned LLM remains intact, allowing for accelerated inference on the deployment stage. We employ PEQA-tuning for task-specific adaptation on LLMs with up to 65 billion parameters. To assess the logical reasoning and language comprehension of PEQA-tuned LLMs, we fine-tune low-bit quantized LLMs using a instruction dataset. Our results show that even when LLMs are quantized to below 4-bit precision, their capabilities in language modeling, few-shot in-context learning, and comprehension can be resiliently restored to (or even improved over) their full-precision original performances with PEQA.
Deploying deep neural networks (DNNs) on microcontrollers (TinyML) is a common trend to process the increasing amount of sensor data generated at the edge, but in practice, resource and latency constraints make it difficult to find optimal DNN candidates. Neural architecture search (NAS) is an excellent approach to automate this search and can easily be combined with DNN compression techniques commonly used in TinyML. However, many NAS techniques are not only computationally expensive, especially hyperparameter optimization (HPO), but also often focus on optimizing only a single objective, e.g., maximizing accuracy, without considering additional objectives such as memory requirements or computational complexity of a DNN, which are key to making deployment at the edge feasible. In this paper, we propose a novel NAS strategy for TinyML based on multi-objective Bayesian optimization (MOBOpt) and an ensemble of competing parametric policies trained using Augmented Random Search (ARS) reinforcement learning (RL) agents. Our methodology aims at efficiently finding tradeoffs between a DNN's predictive accuracy, memory requirements on a given target system, and computational complexity. Our experiments show that we consistently outperform existing MOBOpt approaches on different datasets and architectures such as ResNet-18 and MobileNetV3.
Vision transformers (ViTs) have achieved remarkable performance in various computer vision tasks. However, intensive memory and computation requirements impede ViTs from running on resource-constrained edge devices. Due to the non-normally distributed values after Softmax and GeLU, post-training quantization on ViTs results in severe accuracy degradation. Moreover, conventional methods fail to address the high channel-wise variance in LayerNorm. To reduce the quantization loss and improve classification accuracy, we propose a two-scaled post-training quantization scheme for vision transformer (TSPTQ-ViT). We design the value-aware two-scaled scaling factors (V-2SF) specialized for post-Softmax and post-GeLU values, which leverage the bit sparsity in non-normal distribution to save bit-widths. In addition, the outlier-aware two-scaled scaling factors (O-2SF) are introduced to LayerNorm, alleviating the dominant impacts from outlier values. Our experimental results show that the proposed methods reach near-lossless accuracy drops (<0.5%) on the ImageNet classification task under 8-bit fully quantized ViTs.
In this article, we determine a fundamental anatomical modeling parallelism between low-inertia power systems and Bohr's atomic model. The proposed atomic architecture will serve as a microscopic building block, where we validate the structural analogy of low-inertia power systems using semi-classical quantum approximations in IEEE 9-bus & 39-bus systems. As a future scope of work, detailed modeling & system stability will be investigated by using pre-quantization and geometric quantization methods.
Large language models (LLMs) have revolutionized the field of AI, demonstrating unprecedented capacity across various tasks. However, the inference process for LLMs comes with significant computational costs. In this paper, we propose an efficient LLM inference pipeline that harnesses the power of LLMs. Our approach begins by tapping into the potential of LLMs to accurately perceive and predict the response length with minimal overhead. By leveraging this information, we introduce an efficient sequence scheduling technique that groups queries with similar response lengths into micro-batches. We evaluate our approach on real-world instruction datasets using the LLaMA-based model, and our results demonstrate an impressive 86% improvement in inference throughput without compromising effectiveness. Notably, our method is orthogonal to other inference acceleration techniques, making it a valuable addition to many existing toolkits (e.g., FlashAttention, Quantization) for LLM inference.
Permanent faults induced due to imperfections in the manufacturing process of Deep Neural Network (DNN) accelerators are a major concern, as they negatively impact the manufacturing yield of the chip fabrication process. Fault-aware training is the state-of-the-art approach for mitigating such faults. However, it incurs huge retraining overheads, specifically when used for large DNNs trained on complex datasets. To address this issue, we propose a novel Fault-Aware Quantization (FAQ) technique for mitigating the effects of stuck-at permanent faults in the on-chip weight memory of DNN accelerators at a negligible overhead cost compared to fault-aware retraining while offering comparable accuracy results. We propose a lookup table-based algorithm to achieve ultra-low model conversion time. We present extensive evaluation of the proposed approach using five different DNNs, i.e., ResNet-18, VGG11, VGG16, AlexNet and MobileNetV2, and three different datasets, i.e., CIFAR-10, CIFAR-100 and ImageNet. The results demonstrate that FAQ helps in maintaining the baseline accuracy of the DNNs at low and moderate fault rates without involving costly fault-aware training. For example, for ResNet-18 trained on the CIFAR-10 dataset, at 0.04 fault rate FAQ offers (on average) an increase of 76.38% in accuracy. Similarly, for VGG11 trained on the CIFAR-10 dataset, at 0.04 fault rate FAQ offers (on average) an increase of 70.47% in accuracy. The results also show that FAQ incurs negligible overheads, i.e., less than 5% of the time required to run 1 epoch of retraining. We additionally demonstrate the efficacy of our technique when used in conjunction with fault-aware retraining and show that the use of FAQ inside fault-aware retraining enables fast accuracy recovery.
The excellent performance of deep neural networks is usually accompanied by a large number of parameters and computations, which have limited their usage on the resource-limited edge devices. To address this issue, abundant methods such as pruning, quantization and knowledge distillation have been proposed to compress neural networks and achieved significant breakthroughs. However, most of these compression methods focus on the architecture or the training method of neural networks but ignore the influence from data augmentation. In this paper, we revisit the usage of data augmentation in model compression and give a comprehensive study on the relation between model sizes and their optimal data augmentation policy. To sum up, we mainly have the following three observations: (A) Models in different sizes prefer data augmentation with different magnitudes. Hence, in iterative pruning, data augmentation with varying magnitudes leads to better performance than data augmentation with a consistent magnitude. (B) Data augmentation with a high magnitude may significantly improve the performance of large models but harm the performance of small models. Fortunately, small models can still benefit from strong data augmentations by firstly learning them with "additional parameters" and then discard these "additional parameters" during inference. (C) The prediction of a pre-trained large model can be utilized to measure the difficulty of data augmentation. Thus it can be utilized as a criterion to design better data augmentation policies. We hope this paper may promote more research on the usage of data augmentation in model compression.
Vision transformers (ViTs) quantization offers a promising prospect to facilitate deploying large pre-trained networks on resource-limited devices. Fully-binarized ViTs (Bi-ViT) that pushes the quantization of ViTs to its limit remain largely unexplored and a very challenging task yet, due to their unacceptable performance. Through extensive empirical analyses, we identify the severe drop in ViT binarization is caused by attention distortion in self-attention, which technically stems from the gradient vanishing and ranking disorder. To address these issues, we first introduce a learnable scaling factor to reactivate the vanished gradients and illustrate its effectiveness through theoretical and experimental analyses. We then propose a ranking-aware distillation method to rectify the disordered ranking in a teacher-student framework. Bi-ViT achieves significant improvements over popular DeiT and Swin backbones in terms of Top-1 accuracy and FLOPs. For example, with DeiT-Tiny and Swin-Tiny, our method significantly outperforms baselines by 22.1% and 21.4% respectively, while 61.5x and 56.1x theoretical acceleration in terms of FLOPs compared with real-valued counterparts on ImageNet.
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.
In this paper, we propose a two-bit reconfigurable intelligent surface (RIS)-aided communication system, which mainly consists of a two-bit RIS, a transmitter and a receiver. A corresponding prototype verification system is designed to perform experimental tests in practical environments. The carrier frequency is set as 3.5GHz, and the RIS array possesses 256 units, each of which adopts two-bit phase quantization. In particular, we adopt a self-developed broadband intelligent communication system 40MHz-Net (BICT-40N) terminal in order to fully acquire the channel information. The terminal mainly includes a baseband board and a radio frequency (RF) front-end board, where the latter can achieve 26 dB transmitting link gain and 33 dB receiving link gain. The orthogonal frequency division multiplexing (OFDM) signal is used for the terminal, where the bandwidth is 40MHz and the subcarrier spacing is 625KHz. Also, the terminal supports a series of modulation modes, including QPSK, QAM, etc.Through experimental tests, we validate a few functions and properties of the RIS as follows. First, we validate a novel RIS power consumption model, which considers both the static and the dynamic power consumption. Besides, we demonstrate the existence of the imaging interference and find that two-bit RIS can lower the imaging interference about 10 dBm. Moreover, we verify that the RIS can outperform the metal plate in terms of the beam focusing performance. In addition, we find that the RIS has the ability to improve the channel stationarity. Then, we realize the multi-beam reflection of the RIS utilizing the pattern addition (PA) algorithm. Lastly, we validate the existence of the mutual coupling between different RIS units.
While the numerous parameters in Large Language Models (LLMs) contribute to their superior performance, this massive scale makes them inefficient and memory-hungry. Thus, they are hard to deploy on commodity hardware, such as one single GPU. Given the memory and power constraints of such devices, model compression methods are widely employed to reduce both the model size and inference latency, which essentially trades off model quality in return for improved efficiency. Thus, optimizing this accuracy-efficiency trade-off is crucial for the LLM deployment on commodity hardware. In this paper, we introduce a new perspective to optimize this trade-off by prompting compressed models. Specifically, we first observe that for certain questions, the generation quality of a compressed LLM can be significantly improved by adding carefully designed hard prompts, though this isn't the case for all questions. Based on this observation, we propose a soft prompt learning method where we expose the compressed model to the prompt learning process, aiming to enhance the performance of prompts. Our experimental analysis suggests our soft prompt strategy greatly improves the performance of the 8x compressed LLaMA-7B model (with a joint 4-bit quantization and 50% weight pruning compression), allowing them to match their uncompressed counterparts on popular benchmarks. Also, we demonstrate that these learned prompts can be transferred across various datasets, tasks, and compression levels. Hence with this transferability, we can stitch the soft prompt to a newly compressed model to improve the test-time accuracy in an ``in-situ'' way.
Existing vector quantization (VQ) based autoregressive models follow a two-stage generation paradigm that first learns a codebook to encode images as discrete codes, and then completes generation based on the learned codebook. However, they encode fixed-size image regions into fixed-length codes and ignore their naturally different information densities, which results in insufficiency in important regions and redundancy in unimportant ones, and finally degrades the generation quality and speed. Moreover, the fixed-length coding leads to an unnatural raster-scan autoregressive generation. To address the problem, we propose a novel two-stage framework: (1) Dynamic-Quantization VAE (DQ-VAE) which encodes image regions into variable-length codes based on their information densities for an accurate and compact code representation. (2) DQ-Transformer which thereby generates images autoregressively from coarse-grained (smooth regions with fewer codes) to fine-grained (details regions with more codes) by modeling the position and content of codes in each granularity alternately, through a novel stacked-transformer architecture and shared-content, non-shared position input layers designs. Comprehensive experiments on various generation tasks validate our superiorities in both effectiveness and efficiency. Code will be released at https://github.com/CrossmodalGroup/DynamicVectorQuantization.
The task of table summarization involves generating text that both succinctly and accurately represents the table or a specific set of highlighted cells within a table. While significant progress has been made in table to text generation techniques, models still mostly generate descriptive summaries, which reiterates the information contained within the table in sentences. Through analysis of popular table to text benchmarks (ToTTo (Parikh et al., 2020 and InfoTabs (Gupta et al., 2020) we observe that in order to generate the ideal summary, multiple types of reasoning is needed coupled with access to knowledge beyond the scope of the table. To address this gap, we propose ReTAG, a table and reasoning aware model that uses vector-quantization to infuse different types of analytical reasoning into the output. ReTAG achieves 2.2%, 2.9% improvement on the PARENT metric in the relevant slice of ToTTo and InfoTabs for the table to text generation task over state of the art baselines. Through human evaluation, we observe that output from ReTAG is upto 12% more faithful and analytical compared to a strong table-aware model. To the best of our knowledge, ReTAG is the first model that can controllably use multiple reasoning methods within a structure-aware sequence to sequence model to surpass state of the art performance in multiple table to text tasks. We extend (and open source 35.6K analytical, 55.9k descriptive instances) the ToTTo, InfoTabs datasets with the reasoning categories used in each reference sentences.
As a popular multilingual and multitask pre-trained speech model, Whisper has the problem of curse of multilinguality. To enhance multilingual capabilities in small Whisper models, we propose DQ-Whisper, a novel joint distillation and quantization framework to compress Whisper for efficient inference. Firstly, we propose a novel dynamic matching distillation strategy. Then, a quantization-aware distillation framework is introduced to integrate quantization with distillation. Experimental results on various multilingual datasets show that our suggested distillation approach can effectively enhance the multilingual capabilities of small Whisper models without increasing computational costs. Up to 5.18x reduction in model size is achieved with marginal performance degradation. In addition, quantization is compatible with distillation, which can result in a higher compression rate.
Edge networks call for communication efficient (low overhead) and robust distributed optimization (DO) algorithms. These are, in fact, desirable qualities for DO frameworks, such as federated edge learning techniques, in the presence of data and system heterogeneity, and in scenarios where internode communication is the main bottleneck. Although computationally demanding, Newton-type (NT) methods have been recently advocated as enablers of robust convergence rates in challenging DO problems where edge devices have sufficient computational power. Along these lines, in this work we propose Q-SHED, an original NT algorithm for DO featuring a novel bit-allocation scheme based on incremental Hessian eigenvectors quantization. The proposed technique is integrated with the recent SHED algorithm, from which it inherits appealing features like the small number of required Hessian computations, while being bandwidth-versatile at a bit-resolution level. Our empirical evaluation against competing approaches shows that Q-SHED can reduce by up to 60% the number of communication rounds required for convergence.
The transformer extends its success from the language to the vision domain. Because of the stacked self-attention and cross-attention blocks, the acceleration deployment of vision transformer on GPU hardware is challenging and also rarely studied. This paper thoroughly designs a compression scheme to maximally utilize the GPU-friendly 2:4 fine-grained structured sparsity and quantization. Specially, an original large model with dense weight parameters is first pruned into a sparse one by 2:4 structured pruning, which considers the GPU's acceleration of 2:4 structured sparse pattern with FP16 data type, then the floating-point sparse model is further quantized into a fixed-point one by sparse-distillation-aware quantization aware training, which considers GPU can provide an extra speedup of 2:4 sparse calculation with integer tensors. A mixed-strategy knowledge distillation is used during the pruning and quantization process. The proposed compression scheme is flexible to support supervised and unsupervised learning styles. Experiment results show GPUSQ-ViT scheme achieves state-of-the-art compression by reducing vision transformer models 6.4-12.7 times on model size and 30.3-62 times on FLOPs with negligible accuracy degradation on ImageNet classification, COCO detection and ADE20K segmentation benchmarking tasks. Moreover, GPUSQ-ViT can boost actual deployment performance by 1.39-1.79 times and 3.22-3.43 times of latency and throughput on A100 GPU, and 1.57-1.69 times and 2.11-2.51 times improvement of latency and throughput on AGX Orin.
Speech-driven gesture generation is highly challenging due to the random jitters of human motion. In addition, there is an inherent asynchronous relationship between human speech and gestures. To tackle these challenges, we introduce a novel quantization-based and phase-guided motion-matching framework. Specifically, we first present a gesture VQ-VAE module to learn a codebook to summarize meaningful gesture units. With each code representing a unique gesture, random jittering problems are alleviated effectively. We then use Levenshtein distance to align diverse gestures with different speech. Levenshtein distance based on audio quantization as a similarity metric of corresponding speech of gestures helps match more appropriate gestures with speech, and solves the alignment problem of speech and gestures well. Moreover, we introduce phase to guide the optimal gesture matching based on the semantics of context or rhythm of audio. Phase guides when text-based or speech-based gestures should be performed to make the generated gestures more natural. Extensive experiments show that our method outperforms recent approaches on speech-driven gesture generation. Our code, database, pre-trained models, and demos are available at https://github.com/YoungSeng/QPGesture.
We propose Multiplier-less INTeger (MINT) quantization, a uniform quantization scheme that efficiently compresses weights and membrane potentials in spiking neural networks (SNNs). Unlike previous SNN quantization methods, MINT quantizes memory-intensive membrane potentials to an extremely low precision (2-bit), significantly reducing the memory footprint. MINT also shares the quantization scaling factor between weights and membrane potentials, eliminating the need for multipliers required in conventional uniform quantization. Experimental results show that our method matches the accuracy of full-precision models and other state-of-the-art SNN quantization techniques while surpassing them in memory footprint reduction and hardware cost efficiency at deployment. For example, 2-bit MINT VGG-16 achieves 90.6% accuracy on CIFAR-10, with roughly 93.8% reduction in memory footprint from the full-precision model and 90% reduction in computation energy compared to vanilla uniform quantization at deployment. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/MINT-Quantization.
Isolated training with Gaussian priors (TGP) of the component autoencoders of turbo-autoencoder architectures enables faster, more consistent training and better generalization to arbitrary decoding iterations than training based on deep unfolding. We propose fitting the components via extrinsic information transfer (EXIT) charts to a desired behavior which enables scaling to larger message lengths ($k \approx 1000$) while retaining competitive performance. To the best of our knowledge, this is the first autoencoder that performs close to classical codes in this regime. Although the binary cross-entropy (BCE) loss function optimizes the bit error rate (BER) of the components, the design via EXIT charts enables to focus on the block error rate (BLER). In serially concatenated systems the component-wise TGP approach is well known for inner components with a fixed outer binary interface, e.g., a learned inner code or equalizer, with an outer binary error correcting code. In this paper we extend the component training to structures with an inner and outer autoencoder, where we propose a new 1-bit quantization strategy for the encoder outputs based on the underlying communication problem. Finally, we discuss the model complexity of the learned components during design time (training) and inference and show that the number of weights in the encoder can be reduced by 99.96 %.
In spiking neural networks (SNN), at each node, an incoming sequence of weighted Dirac pulses is converted into an output sequence of weighted Dirac pulses by a leaky-integrate-and-fire (LIF) neuron model based on spike aggregation and thresholding. We show that this mapping can be understood as a quantization operator and state a corresponding formula for the quantization error by means of the Alexiewicz norm. This analysis has implications for rethinking re-initialization in the LIF model, leading to the proposal of 'reset-to-mod' as a modulo-based reset variant.