Graph Neural Network (GNN) ushered in a new era of machine learning with interconnected datasets. While traditional neural networks can only be trained on independent samples, GNN allows for the inclusion of inter-sample interactions in the training process. This gain, however, incurs additional memory cost, rendering most GNNs unscalable for real-world applications involving vast and complicated networks with tens of millions of nodes (e.g., social circles, web graphs, and brain graphs). This means that storing the graph in the main memory can be difficult, let alone training the GNN model with significantly less GPU memory. While much of the recent literature has focused on either mini-batching GNN methods or quantization, graph reduction methods remain largely scarce. Furthermore, present graph reduction approaches have several drawbacks. First, most graph reduction focuses only on the inference stage (e.g., condensation and distillation) and requires full graph GNN training, which does not reduce training memory footprint. Second, many methods focus solely on the graph's structural aspect, ignoring the initial population feature-label distribution, resulting in a skewed post-reduction label distribution. Here, we propose a Feature-Label COnstrained graph Net collapse, FALCON, to address these limitations. Our three core contributions lie in (i) designing FALCON, a topology-aware graph reduction technique that preserves feature-label distribution; (ii) implementation of FALCON with other memory reduction methods (i.e., mini-batched GNN and quantization) for further memory reduction; (iii) extensive benchmarking and ablation studies against SOTA methods to evaluate FALCON memory reduction. Our extensive results show that FALCON can significantly collapse various public datasets while achieving equal prediction quality across GNN models. Code: https://github.com/basiralab/FALCON
Effective communication protocols in multi-agent reinforcement learning (MARL) are critical to fostering cooperation and enhancing team performance. To leverage communication, many previous works have proposed to compress local information into a single message and broadcast it to all reachable agents. This simplistic messaging mechanism, however, may fail to provide adequate, critical, and relevant information to individual agents, especially in severely bandwidth-limited scenarios. This motivates us to develop context-aware communication schemes for MARL, aiming to deliver personalized messages to different agents. Our communication protocol, named CACOM, consists of two stages. In the first stage, agents exchange coarse representations in a broadcast fashion, providing context for the second stage. Following this, agents utilize attention mechanisms in the second stage to selectively generate messages personalized for the receivers. Furthermore, we employ the learned step size quantization (LSQ) technique for message quantization to reduce the communication overhead. To evaluate the effectiveness of CACOM, we integrate it with both actor-critic and value-based MARL algorithms. Empirical results on cooperative benchmark tasks demonstrate that CACOM provides evident performance gains over baselines under communication-constrained scenarios. The code is publicly available at https://github.com/LXXXXR/CACOM.
Deep neural networks are characterized by multiple symmetrical, equi-loss solutions that are redundant. Thus, the order of neurons in a layer and feature maps can be given arbitrary permutations, without affecting (or minimally affecting) their output. If we shuffle these neurons, or if we apply to them some perturbations (like fine-tuning) can we put them back in the original order i.e. re-synchronize? Is there a possible corruption threat? Answering these questions is important for applications like neural network white-box watermarking for ownership tracking and integrity verification. We advance a method to re-synchronize the order of permuted neurons. Our method is also effective if neurons are further altered by parameter pruning, quantization, and fine-tuning, showing robustness to integrity attacks. Additionally, we provide theoretical and practical evidence for the usual means to corrupt the integrity of the model, resulting in a solution to counter it. We test our approach on popular computer vision datasets and models, and we illustrate the threat and our countermeasure on a popular white-box watermarking method.
Motivated by high resource costs of centralized machine learning schemes as well as data privacy concerns, federated learning (FL) emerged as an efficient alternative that relies on aggregating locally trained models rather than collecting clients' potentially private data. In practice, available resources and data distributions vary from one client to another, creating an inherent system heterogeneity that leads to deterioration of the performance of conventional FL algorithms. In this work, we present a federated quantization-based self-supervised learning scheme (Fed-QSSL) designed to address heterogeneity in FL systems. At clients' side, to tackle data heterogeneity we leverage distributed self-supervised learning while utilizing low-bit quantization to satisfy constraints imposed by local infrastructure and limited communication resources. At server's side, Fed-QSSL deploys de-quantization, weighted aggregation and re-quantization, ultimately creating models personalized to both data distribution as well as specific infrastructure of each client's device. We validated the proposed algorithm on real world datasets, demonstrating its efficacy, and theoretically analyzed impact of low-bit training on the convergence and robustness of the learned models.
3D Gaussian Splatting has recently emerged as a highly promising technique for modeling of static 3D scenes. In contrast to Neural Radiance Fields, it utilizes efficient rasterization allowing for very fast rendering at high-quality. However, the storage size is significantly higher, which hinders practical deployment, e.g. on resource constrained devices. In this paper, we introduce a compact scene representation organizing the parameters of 3D Gaussian Splatting (3DGS) into a 2D grid with local homogeneity, ensuring a drastic reduction in storage requirements without compromising visual quality during rendering. Central to our idea is the explicit exploitation of perceptual redundancies present in natural scenes. In essence, the inherent nature of a scene allows for numerous permutations of Gaussian parameters to equivalently represent it. To this end, we propose a novel highly parallel algorithm that regularly arranges the high-dimensional Gaussian parameters into a 2D grid while preserving their neighborhood structure. During training, we further enforce local smoothness between the sorted parameters in the grid. The uncompressed Gaussians use the same structure as 3DGS, ensuring a seamless integration with established renderers. Our method achieves a reduction factor of 17x to 42x in size for complex scenes with no increase in training time, marking a substantial leap forward in the domain of 3D scene distribution and consumption. Additional information can be found on our project page: https://fraunhoferhhi.github.io/Self-Organizing-Gaussians/
Recently segment anything model (SAM) has shown powerful segmentation capability and has drawn great attention in computer vision fields. Massive following works have developed various applications based on the pre-trained SAM and achieved impressive performance on downstream vision tasks. However, SAM consists of heavy architectures and requires massive computational capacity, which hinders the further application of SAM on computation constrained edge devices. To this end, in this paper we propose a framework to obtain a tiny segment anything model (TinySAM) while maintaining the strong zero-shot performance. We first propose a full-stage knowledge distillation method with hard prompt sampling and hard mask weighting strategy to distill a lightweight student model. We also adapt the post-training quantization to the prompt-based segmentation task and further reduce the computational cost. Moreover, a hierarchical segmenting everything strategy is proposed to accelerate the everything inference by $2\times$ with almost no performance degradation. With all these proposed methods, our TinySAM leads to orders of magnitude computational reduction and pushes the envelope for efficient segment anything task. Extensive experiments on various zero-shot transfer tasks demonstrate the significantly advantageous performance of our TinySAM against counterpart methods. Codes are available at https://github.com/xinghaochen/TinySAM and https://gitee.com/mindspore/models/tree/master/research/cv/TinySAM.
Quantization replaces floating point arithmetic with integer arithmetic in deep neural network models, providing more efficient on-device inference with less power and memory. In this work, we propose a framework for formally verifying properties of quantized neural networks. Our baseline technique is based on integer linear programming which guarantees both soundness and completeness. We then show how efficiency can be improved by utilizing gradient-based heuristic search methods and also bound-propagation techniques. We evaluate our approach on perception networks quantized with PyTorch. Our results show that we can verify quantized networks with better scalability and efficiency than the previous state of the art.
Style transfer for out-of-domain (OOD) singing voice synthesis (SVS) focuses on generating high-quality singing voices with unseen styles (such as timbre, emotion, pronunciation, and articulation skills) derived from reference singing voice samples. However, the endeavor to model the intricate nuances of singing voice styles is an arduous task, as singing voices possess a remarkable degree of expressiveness. Moreover, existing SVS methods encounter a decline in the quality of synthesized singing voices in OOD scenarios, as they rest upon the assumption that the target vocal attributes are discernible during the training phase. To overcome these challenges, we propose StyleSinger, the first singing voice synthesis model for zero-shot style transfer of out-of-domain reference singing voice samples. StyleSinger incorporates two critical approaches for enhanced effectiveness: 1) the Residual Style Adaptor (RSA) which employs a residual quantization module to capture diverse style characteristics in singing voices, and 2) the Uncertainty Modeling Layer Normalization (UMLN) to perturb the style attributes within the content representation during the training phase and thus improve the model generalization. Our extensive evaluations in zero-shot style transfer undeniably establish that StyleSinger outperforms baseline models in both audio quality and similarity to the reference singing voice samples. Access to singing voice samples can be found at https://aaronz345.github.io/StyleSingerDemo/.
While transformer models have been highly successful, they are computationally inefficient. We observe that for each layer, the full width of the layer may be needed only for a small subset of tokens inside a batch and that the "effective" width needed to process a token can vary from layer to layer. Motivated by this observation, we introduce the Adaptive Computation Module (ACM), a generic module that dynamically adapts its computational load to match the estimated difficulty of the input on a per-token basis. An ACM consists of a sequence of learners that progressively refine the output of their preceding counterparts. An additional gating mechanism determines the optimal number of learners to execute for each token. We also propose a distillation technique to replace any pre-trained model with an "ACMized" variant. Our evaluation of transformer models in computer vision and speech recognition demonstrates that substituting layers with ACMs significantly reduces inference costs without degrading the downstream accuracy for a wide interval of user-defined budgets.
Transformer-based large language models (e.g., BERT and GPT) achieve great success, and fine-tuning, which tunes a pre-trained model on a task-specific dataset, is the standard practice to utilize these models for downstream tasks. However, Transformer fine-tuning has long running time and high memory consumption due to the large size of the models. We propose the SPT system to fine-tune Transformer-based models efficiently by introducing sparsity. We observe that the memory consumption of Transformer mainly comes from storing attention weights for multi-head attention (MHA), and the majority of running time is spent on feed-forward network (FFN). Thus, we design the sparse MHA module, which computes and stores only large attention weights to reduce memory consumption, and the routed FFN module, which dynamically activates a subset of model parameters for each token to reduce computation cost. We implement SPT on PyTorch and customize CUDA kernels to run sparse MHA and routed FFN efficiently. Specifically, we use product quantization to identify the large attention weights and compute attention via sparse matrix multiplication for sparse MHA. For routed FFN, we batch the tokens according to their activated model parameters for efficient computation. We conduct extensive experiments to evaluate SPT on various model configurations. The results show that SPT consistently outperforms well-optimized baselines, reducing the peak memory consumption by up to 50% and accelerating fine-tuning by up to 2.2x.
Although neural networks have made remarkable advancements in various applications, they require substantial computational and memory resources. Network quantization is a powerful technique to compress neural networks, allowing for more efficient and scalable AI deployments. Recently, Re-parameterization has emerged as a promising technique to enhance model performance while simultaneously alleviating the computational burden in various computer vision tasks. However, the accuracy drops significantly when applying quantization on the re-parameterized networks. We identify that the primary challenge arises from the large variation in weight distribution across the original branches. To address this issue, we propose a coarse & fine weight splitting (CFWS) method to reduce quantization error of weight, and develop an improved KL metric to determine optimal quantization scales for activation. To the best of our knowledge, our approach is the first work that enables post-training quantization applicable on re-parameterized networks. For example, the quantized RepVGG-A1 model exhibits a mere 0.3% accuracy loss. The code is in https://github.com/NeonHo/Coarse-Fine-Weight-Split.git
This study examines 4-bit quantization methods like GPTQ in large language models (LLMs), highlighting GPTQ's overfitting and limited enhancement in Zero-Shot tasks. While prior works merely focusing on zero-shot measurement, we extend task scope to more generative categories such as code generation and abstractive summarization, in which we found that INT4 quantization can significantly underperform. However, simply shifting to higher precision formats like FP6 has been particularly challenging, thus overlooked, due to poor performance caused by the lack of sophisticated integration and system acceleration strategies on current AI hardware. Our results show that FP6, even with a coarse-grain quantization scheme, performs robustly across various algorithms and tasks, demonstrating its superiority in accuracy and versatility. Notably, with the FP6 quantization, \codestar-15B model performs comparably to its FP16 counterpart in code generation, and for smaller models like the 406M it closely matches their baselines in summarization. Neither can be achieved by INT4. To better accommodate various AI hardware and achieve the best system performance, we propose a novel 4+2 design for FP6 to achieve similar latency to the state-of-the-art INT4 fine-grain quantization. With our design, FP6 can become a promising solution to the current 4-bit quantization methods used in LLMs.
Large Language Models (LLMs) trained with self-supervision on vast corpora of web text fit to the social biases of that text. Without intervention, these social biases persist in the model's predictions in downstream tasks, leading to representational harm. Many strategies have been proposed to mitigate the effects of inappropriate social biases learned during pretraining. Simultaneously, methods for model compression have become increasingly popular to reduce the computational burden of LLMs. Despite the popularity and need for both approaches, little work has been done to explore the interplay between these two. We perform a carefully controlled study of the impact of model compression via quantization and knowledge distillation on measures of social bias in LLMs. Longer pretraining and larger models led to higher social bias, and quantization showed a regularizer effect with its best trade-off around 20% of the original pretraining time.
Offline meta-reinforcement learning (meta-RL) methods, which adapt to unseen target tasks with prior experience, are essential in robot control tasks. Current methods typically utilize task contexts and skills as prior experience, where task contexts are related to the information within each task and skills represent a set of temporally extended actions for solving subtasks. However, these methods still suffer from limited performance when adapting to unseen target tasks, mainly because the learned prior experience lacks generalization, i.e., they are unable to extract effective prior experience from meta-training tasks by exploration and learning of continuous latent spaces. We propose a framework called decoupled meta-reinforcement learning (DCMRL), which (1) contrastively restricts the learning of task contexts through pulling in similar task contexts within the same task and pushing away different task contexts of different tasks, and (2) utilizes a Gaussian quantization variational autoencoder (GQ-VAE) for clustering the Gaussian distributions of the task contexts and skills respectively, and decoupling the exploration and learning processes of their spaces. These cluster centers which serve as representative and discrete distributions of task context and skill are stored in task context codebook and skill codebook, respectively. DCMRL can acquire generalizable prior experience and achieve effective adaptation to unseen target tasks during the meta-testing phase. Experiments in the navigation and robot manipulation continuous control tasks show that DCMRL is more effective than previous meta-RL methods with more generalizable prior experience.
Large Language Models (LLMs) stand out for their impressive performance in intricate language modeling tasks. However, their demanding computational and memory needs pose obstacles for broad use on edge devices. Quantization is then introduced to boost LLMs' on-device efficiency. Recent works show that 8-bit or lower weight quantization is feasible with minimal impact on end-to-end task performance, while the activation is still not quantized. On the other hand, mainstream commodity edge devices still struggle to execute these sub-8-bit quantized networks effectively. In this paper, we propose Agile-Quant, an activation-guided quantization framework for popular Large Language Models (LLMs), and implement an end-to-end accelerator on multiple edge devices for faster inference. Considering the hardware profiling and activation analysis, we first introduce a basic activation quantization strategy to balance the trade-off of task performance and real inference speed. Then we leverage the activation-aware token pruning technique to reduce the outliers and the adverse impact on attentivity. Ultimately, we utilize the SIMD-based 4-bit multiplier and our efficient TRIP matrix multiplication to implement the accelerator for LLMs on the edge. We apply our framework on different scales of LLMs including LLaMA, OPT, and BLOOM with 4-bit or 8-bit for the activation and 4-bit for the weight quantization. Experiments show that Agile-Quant achieves simultaneous quantization of model weights and activations while maintaining task performance comparable to existing weight-only quantization methods. Moreover, in the 8- and 4-bit scenario, Agile-Quant achieves an on-device speedup of up to 2.55x compared to its FP16 counterparts across multiple edge devices, marking a pioneering advancement in this domain. Code: https://github.com/shawnricecake/agile-quant
Large language models (LLMs) have shown remarkable capabilities in various tasks. However their huge model size and the consequent demand for computational and memory resources also pose challenges to model deployment. Currently, 4-bit post-training quantization (PTQ) has achieved some success in LLMs, reducing the memory footprint by approximately 75% compared to FP16 models, albeit with some accuracy loss. In this paper, we propose SmoothQuant+, an accurate and efficient 4-bit weight-only PTQ that requires no additional training, which enables lossless in accuracy for LLMs for the first time. Based on the fact that the loss of weight quantization is amplified by the activation outliers, SmoothQuant+ smoothes the activation outliers by channel before quantization, while adjusting the corresponding weights for mathematical equivalence, and then performs group-wise 4-bit weight quantization for linear layers. We have integrated SmoothQuant+ into the vLLM framework, an advanced high-throughput inference engine specially developed for LLMs, and equipped it with an efficient W4A16 CUDA kernels, so that vLLM can seamlessly support SmoothQuant+ 4-bit weight quantization. Our results show that, with SmoothQuant+, the Code Llama-34B model can be quantized and deployed on a A100 40GB GPU, achieving lossless accuracy and a throughput increase of 1.9 to 4.0 times compared to the FP16 model deployed on two A100 40GB GPUs. Moreover, the latency per token is only 68% of the FP16 model deployed on two A100 40GB GPUs. This is the state-of-the-art 4-bit weight quantization for LLMs as we know.
Spatio-temporal forecasting is crucial in various fields and requires a careful balance between identifying subtle patterns and filtering out noise. Vector quantization (VQ) appears well-suited for this purpose, as it quantizes input vectors into a set of codebook vectors or patterns. Although VQ has shown promise in various computer vision tasks, it surprisingly falls short in enhancing the accuracy of spatio-temporal forecasting. We attribute this to two main issues: inaccurate optimization due to non-differentiability and limited representation power in hard-VQ. To tackle these challenges, we introduce Differentiable Sparse Soft-Vector Quantization (SVQ), the first VQ method to enhance spatio-temporal forecasting. SVQ balances detail preservation with noise reduction, offering full differentiability and a solid foundation in sparse regression. Our approach employs a two-layer MLP and an extensive codebook to streamline the sparse regression process, significantly cutting computational costs while simplifying training and improving performance. Empirical studies on five spatio-temporal benchmark datasets show SVQ achieves state-of-the-art results, including a 7.9% improvement on the WeatherBench-S temperature dataset and an average mean absolute error reduction of 9.4% in video prediction benchmarks (Human3.6M, KTH, and KittiCaltech), along with a 17.3% enhancement in image quality (LPIPS). Code is publicly available at https://github.com/Pachark/SVQ-Forecasting.
In recent years, there has been a significant trend in deep neural networks (DNNs), particularly transformer-based models, of developing ever-larger and more capable models. While they demonstrate state-of-the-art performance, their growing scale requires increased computational resources (e.g., GPUs with greater memory capacity). To address this problem, quantization techniques (i.e., low-bit-precision representation and matrix multiplication) have been proposed. Most quantization techniques employ a static strategy in which the model parameters are quantized, either during training or inference, without considering the test-time sample. In contrast, dynamic quantization techniques, which have become increasingly popular, adapt during inference based on the input provided, while maintaining full-precision performance. However, their dynamic behavior and average-case performance assumption makes them vulnerable to a novel threat vector -- adversarial attacks that target the model's efficiency and availability. In this paper, we present QuantAttack, a novel attack that targets the availability of quantized models, slowing down the inference, and increasing memory usage and energy consumption. We show that carefully crafted adversarial examples, which are designed to exhaust the resources of the operating system, can trigger worst-case performance. In our experiments, we demonstrate the effectiveness of our attack on vision transformers on a wide range of tasks, both uni-modal and multi-modal. We also examine the effect of different attack variants (e.g., a universal perturbation) and the transferability between different models.
Pruning-quantization joint learning always facilitates the deployment of deep neural networks (DNNs) on resource-constrained edge devices. However, most existing methods do not jointly learn a global criterion for pruning and quantization in an interpretable way. In this paper, we propose a novel physics inspired criterion for pruning-quantization joint learning (PIC-PQ), which is explored from an analogy we first draw between elasticity dynamics (ED) and model compression (MC). Specifically, derived from Hooke's law in ED, we establish a linear relationship between the filters' importance distribution and the filter property (FP) by a learnable deformation scale in the physics inspired criterion (PIC). Furthermore, we extend PIC with a relative shift variable for a global view. To ensure feasibility and flexibility, available maximum bitwidth and penalty factor are introduced in quantization bitwidth assignment. Experiments on benchmarks of image classification demonstrate that PIC-PQ yields a good trade-off between accuracy and bit-operations (BOPs) compression ratio e.g., 54.96X BOPs compression ratio in ResNet56 on CIFAR10 with 0.10% accuracy drop and 53.24X in ResNet18 on ImageNet with 0.61% accuracy drop). The code will be available at https://github.com/fanxxxxyi/PIC-PQ.
Compressing large language models (LLMs), often consisting of billions of parameters, provides faster inference, smaller memory footprints, and enables local deployment. Two standard compression techniques are pruning and quantization, with the former eliminating redundant connections in model layers and the latter representing model parameters with fewer bits. The key tradeoff is between the degree of compression and the impact on the quality of the compressed model. Existing research on LLM compression primarily focuses on performance in terms of general metrics like perplexity or downstream task accuracy. More fine-grained metrics, such as those measuring parametric knowledge, remain significantly underexplored. To help bridge this gap, we present a comprehensive analysis across multiple model families (ENCODER, ENCODER-DECODER, and DECODER) using the LAMA and LM-HARNESS benchmarks in order to systematically quantify the effect of commonly employed compression techniques on model performance. A particular focus is on tradeoffs involving parametric knowledge, with the goal of providing practitioners with practical insights to help make informed decisions on compression. We release our codebase1 to enable further research.