Deep learning-based autoencoders have been employed to compress and reconstruct channel state information (CSI) in frequency-division duplex systems. Practical implementations require judicious quantization of encoder outputs for digital transmission. In this paper, we propose a novel quantization module with bit allocation among encoder outputs and develop a method for joint training the module and the autoencoder. To enhance learning performance, we design a loss function that adaptively weights the quantization loss and the logarithm of reconstruction loss. Simulation results show the performance gain of the proposed method over existing baselines.
The rapid rise of Language Models (LMs) has expanded the capabilities of natural language processing, powering applications from text generation to complex decision-making. While state-of-the-art LMs often boast hundreds of billions of parameters and are primarily deployed in data centers, recent trends show a growing focus on compact models-typically under 10 billion parameters-enabled by techniques such as quantization and other model compression techniques. This shift paves the way for LMs on edge devices, offering potential benefits such as enhanced privacy, reduced latency, and improved data sovereignty. However, the inherent complexity of even these smaller models, combined with the limited computing resources of edge hardware, raises critical questions about the practical trade-offs in executing LM inference outside the cloud. To address these challenges, we present a comprehensive evaluation of generative LM inference on representative CPU-based and GPU-accelerated edge devices. Our study measures key performance indicators-including memory usage, inference speed, and energy consumption-across various device configurations. Additionally, we examine throughput-energy trade-offs, cost considerations, and usability, alongside an assessment of qualitative model performance. While quantization helps mitigate memory overhead, it does not fully eliminate resource bottlenecks, especially for larger models. Our findings quantify the memory and energy constraints that must be considered for practical real-world deployments, offering concrete insights into the trade-offs between model size, inference performance, and efficiency. The exploration of LMs at the edge is still in its early stages. We hope this study provides a foundation for future research, guiding the refinement of models, the enhancement of inference efficiency, and the advancement of edge-centric AI systems.
Reducing the memory footprint of Machine Learning (ML) models, especially Deep Neural Networks (DNNs), is imperative to facilitate their deployment on resource-constrained edge devices. However, a notable drawback of DNN models lies in their susceptibility to adversarial attacks, wherein minor input perturbations can deceive them. A primary challenge revolves around the development of accurate, resilient, and compact DNN models suitable for deployment on resource-constrained edge devices. This paper presents the outcomes of a compact DNN model that exhibits resilience against both black-box and white-box adversarial attacks. This work has achieved this resilience through training with the QKeras quantization-aware training framework. The study explores the potential of QKeras and an adversarial robustness technique, Jacobian Regularization (JR), to co-optimize the DNN architecture through per-layer JR methodology. As a result, this paper has devised a DNN model employing this co-optimization strategy based on Stochastic Ternary Quantization (STQ). Its performance was compared against existing DNN models in the face of various white-box and black-box attacks. The experimental findings revealed that, the proposed DNN model had small footprint and on average, it exhibited better performance than Quanos and DS-CNN MLCommons/TinyML (MLC/T) benchmarks when challenged with white-box and black-box attacks, respectively, on the CIFAR-10 image and Google Speech Commands audio datasets.
Despite recent advancements in federated learning (FL), the integration of generative models into FL has been limited due to challenges such as high communication costs and unstable training in heterogeneous data environments. To address these issues, we propose PRISM, a FL framework tailored for generative models that ensures (i) stable performance in heterogeneous data distributions and (ii) resource efficiency in terms of communication cost and final model size. The key of our method is to search for an optimal stochastic binary mask for a random network rather than updating the model weights, identifying a sparse subnetwork with high generative performance; i.e., a ``strong lottery ticket''. By communicating binary masks in a stochastic manner, PRISM minimizes communication overhead. This approach, combined with the utilization of maximum mean discrepancy (MMD) loss and a mask-aware dynamic moving average aggregation method (MADA) on the server side, facilitates stable and strong generative capabilities by mitigating local divergence in FL scenarios. Moreover, thanks to its sparsifying characteristic, PRISM yields a lightweight model without extra pruning or quantization, making it ideal for environments such as edge devices. Experiments on MNIST, FMNIST, CelebA, and CIFAR10 demonstrate that PRISM outperforms existing methods, while maintaining privacy with minimal communication costs. PRISM is the first to successfully generate images under challenging non-IID and privacy-preserving FL environments on complex datasets, where previous methods have struggled.
In the field of emergent language, efforts have traditionally focused on developing communication protocols through interactions between agents in referential games. However, the aspect of internal language learning, where language serves not only as a communicative tool with others but also as a means for individual thinking, self-reflection, and problem-solving remains underexplored. Developing a language through self-play, without another agent's involvement, poses a unique challenge. It requires an agent to craft symbolic representations and train them using direct gradient methods. The challenge here is that if an agent attempts to learn symbolic representations through self-play using conventional modeling and techniques such as REINFORCE, the solution will offer no advantage over previous multi-agent approaches. We introduce VQEL, a novel method that incorporates Vector Quantization into the agents' architecture, enabling them to autonomously invent and develop discrete symbolic representations in a self-play referential game. Following the self-play phase, agents can enhance their language through reinforcement learning and interactions with other agents in the mutual-play phase. Our experiments across various datasets demonstrate that VQEL not only outperforms the traditional REINFORCE method but also benefits from improved control and reduced susceptibility to collapse, thanks to the incorporation of vector quantization.
Implicit Neural Representations (INRs) have emerged as a powerful approach for video representation, offering versatility across tasks such as compression and inpainting. However, their implicit formulation limits both interpretability and efficacy, undermining their practicality as a comprehensive solution. We propose a novel video representation based on deformable 2D Gaussian splatting, dubbed D2GV, which aims to achieve three key objectives: 1) improved efficiency while delivering superior quality; 2) enhanced scalability and interpretability; and 3) increased friendliness for downstream tasks. Specifically, we initially divide the video sequence into fixed-length Groups of Pictures (GoP) to allow parallel training and linear scalability with video length. For each GoP, D2GV represents video frames by applying differentiable rasterization to 2D Gaussians, which are deformed from a canonical space into their corresponding timestamps. Notably, leveraging efficient CUDA-based rasterization, D2GV converges fast and decodes at speeds exceeding 400 FPS, while delivering quality that matches or surpasses state-of-the-art INRs. Moreover, we incorporate a learnable pruning and quantization strategy to streamline D2GV into a more compact representation. We demonstrate D2GV's versatility in tasks including video interpolation, inpainting and denoising, underscoring its potential as a promising solution for video representation. Code is available at: https://github.com/Evan-sudo/D2GV.
One-step diffusion-based image super-resolution (OSDSR) models are showing increasingly superior performance nowadays. However, although their denoising steps are reduced to one and they can be quantized to 8-bit to reduce the costs further, there is still significant potential for OSDSR to quantize to lower bits. To explore more possibilities of quantized OSDSR, we propose an efficient method, Quantization via reverse-module and timestep-retraining for OSDSR, named QArtSR. Firstly, we investigate the influence of timestep value on the performance of quantized models. Then, we propose Timestep Retraining Quantization (TRQ) and Reversed Per-module Quantization (RPQ) strategies to calibrate the quantized model. Meanwhile, we adopt the module and image losses to update all quantized modules. We only update the parameters in quantization finetuning components, excluding the original weights. To ensure that all modules are fully finetuned, we add extended end-to-end training after per-module stage. Our 4-bit and 2-bit quantization experimental results indicate that QArtSR obtains superior effects against the recent leading comparison methods. The performance of 4-bit QArtSR is close to the full-precision one. Our code will be released at https://github.com/libozhu03/QArtSR.
Despite recent efforts in understanding the compression impact on large language models (LLMs) in terms of their downstream task performance and trustworthiness on relatively simpler uni-modal benchmarks (for example, question answering, common sense reasoning), their detailed study on multi-modal Large Vision-Language Models (LVLMs) is yet to be unveiled. Towards mitigating this gap, we present LVLM-Compress-Bench, a framework to first thoroughly study the broad impact of compression on the generative performance of LVLMs with multi-modal input driven tasks. In specific, we consider two major classes of compression for autoregressive models, namely KV cache and weight compression, for the dynamically growing intermediate cache and static weights, respectively. We use four LVLM variants of the popular LLaVA framework to present our analysis via integrating various state-of-the-art KV and weight compression methods including uniform, outlier-reduced, and group quantization for the KV cache and weights. With this framework we demonstrate on ten different multi-modal datasets with different capabilities including recognition, knowledge, language generation, spatial awareness, visual reasoning, hallucination and visual illusion identification, toxicity, stereotypes and bias. In specific, our framework demonstrates the compression impact on both general and ethically critical metrics leveraging a combination of real world and synthetic datasets to encompass diverse societal intersectional attributes. Extensive experimental evaluations yield diverse and intriguing observations on the behavior of LVLMs at different quantization budget of KV and weights, in both maintaining and losing performance as compared to the baseline model with FP16 data format. Code will be open-sourced at https://github.com/opengear-project/LVLM-compress-bench.
Autoregressive (AR) models for image generation typically adopt a two-stage paradigm of vector quantization and raster-scan ``next-token prediction", inspired by its great success in language modeling. However, due to the huge modality gap, image autoregressive models may require a systematic reevaluation from two perspectives: tokenizer format and regression direction. In this paper, we introduce the frequency progressive autoregressive (\textbf{FAR}) paradigm and instantiate FAR with the continuous tokenizer. Specifically, we identify spectral dependency as the desirable regression direction for FAR, wherein higher-frequency components build upon the lower one to progressively construct a complete image. This design seamlessly fits the causality requirement for autoregressive models and preserves the unique spatial locality of image data. Besides, we delve into the integration of FAR and the continuous tokenizer, introducing a series of techniques to address optimization challenges and improve the efficiency of training and inference processes. We demonstrate the efficacy of FAR through comprehensive experiments on the ImageNet dataset and verify its potential on text-to-image generation.
Quantized Neural Networks (QNNs) have emerged as a promising solution for reducing model size and computational costs, making them well-suited for deployment in edge and resource-constrained environments. While quantization is known to disrupt gradient propagation and enhance robustness against pixel-level adversarial attacks, its effectiveness against patch-based adversarial attacks remains largely unexplored. In this work, we demonstrate that adversarial patches remain highly transferable across quantized models, achieving over 70\% attack success rates (ASR) even at extreme bit-width reductions (e.g., 2-bit). This challenges the common assumption that quantization inherently mitigates adversarial threats. To address this, we propose Quantization-Aware Defense Training with Randomization (QADT-R), a novel defense strategy that integrates Adaptive Quantization-Aware Patch Generation (A-QAPA), Dynamic Bit-Width Training (DBWT), and Gradient-Inconsistent Regularization (GIR) to enhance resilience against highly transferable patch-based attacks. A-QAPA generates adversarial patches within quantized models, ensuring robustness across different bit-widths. DBWT introduces bit-width cycling during training to prevent overfitting to a specific quantization setting, while GIR injects controlled gradient perturbations to disrupt adversarial optimization. Extensive evaluations on CIFAR-10 and ImageNet show that QADT-R reduces ASR by up to 25\% compared to prior defenses such as PBAT and DWQ. Our findings further reveal that PBAT-trained models, while effective against seen patch configurations, fail to generalize to unseen patches due to quantization shift. Additionally, our empirical analysis of gradient alignment, spatial sensitivity, and patch visibility provides insights into the mechanisms that contribute to the high transferability of patch-based attacks in QNNs.
Diffusion Transformer (DiT) has now become the preferred choice for building image generation models due to its great generation capability. Unlike previous convolution-based UNet models, DiT is purely composed of a stack of transformer blocks, which renders DiT excellent in scalability like large language models. However, the growing model size and multi-step sampling paradigm bring about considerable pressure on deployment and inference. In this work, we propose a post-training quantization framework tailored for Diffusion Transforms to tackle these challenges. We firstly locate that the quantization difficulty of DiT mainly originates from the time-dependent channel-specific outliers. We propose a timestep-aware shift-and-scale strategy to smooth the activation distribution to reduce the quantization error. Secondly, based on the observation that activations of adjacent timesteps have similar distributions, we utilize a hierarchical clustering scheme to divide the denoising timesteps into multiple groups. We further design a re-parameterization scheme which absorbs the quantization parameters into nearby module to avoid redundant computations. Comprehensive experiments demonstrate that out PTQ method successfully quantize the Diffusion Transformer into 8-bit weight and 8-bit activation (W8A8) with state-of-the-art FiD score. And our method can further quantize DiT model into 4-bit weight and 8-bit activation (W4A8) without sacrificing generation quality.
Diffusion models have been widely adopted in image and video generation. However, their complex network architecture leads to high inference overhead for its generation process. Existing diffusion quantization methods primarily focus on the quantization of the model structure while ignoring the impact of time-steps variation during sampling. At the same time, most current approaches fail to account for significant activations that cannot be eliminated, resulting in substantial performance degradation after quantization. To address these issues, we propose Time-Rotation Diffusion Quantization (TR-DQ), a novel quantization method incorporating time-step and rotation-based optimization. TR-DQ first divides the sampling process based on time-steps and applies a rotation matrix to smooth activations and weights dynamically. For different time-steps, a dedicated hyperparameter is introduced for adaptive timing modeling, which enables dynamic quantization across different time steps. Additionally, we also explore the compression potential of Classifier-Free Guidance (CFG-wise) to establish a foundation for subsequent work. TR-DQ achieves state-of-the-art (SOTA) performance on image generation and video generation tasks and a 1.38-1.89x speedup and 1.97-2.58x memory reduction in inference compared to existing quantization methods.
Recently, Diffusion Transformers (DiTs) have emerged as a dominant architecture in video generation, surpassing U-Net-based models in terms of performance. However, the enhanced capabilities of DiTs come with significant drawbacks, including increased computational and memory costs, which hinder their deployment on resource-constrained devices. Current acceleration techniques, such as quantization and cache mechanism, offer limited speedup and are often applied in isolation, failing to fully address the complexities of DiT architectures. In this paper, we propose QuantCache, a novel training-free inference acceleration framework that jointly optimizes hierarchical latent caching, adaptive importance-guided quantization, and structural redundancy-aware pruning. QuantCache achieves an end-to-end latency speedup of 6.72$\times$ on Open-Sora with minimal loss in generation quality. Extensive experiments across multiple video generation benchmarks demonstrate the effectiveness of our method, setting a new standard for efficient DiT inference. The code and models will be available at https://github.com/JunyiWuCode/QuantCache.
Segment Anything Model (SAM) exhibits remarkable zero-shot segmentation capability; however, its prohibitive computational costs make edge deployment challenging. Although post-training quantization (PTQ) offers a promising compression solution, existing methods yield unsatisfactory results when applied to SAM, owing to its specialized model components and promptable workflow: (i) The mask decoder's attention exhibits extreme outliers, and we find that aggressive clipping (ranging down to even 100$\times$), instead of smoothing or isolation, is effective in suppressing outliers while maintaining semantic capabilities. Unfortunately, traditional metrics (e.g., MSE) fail to provide such large-scale clipping. (ii) Existing reconstruction methods potentially neglect prompts' intention, resulting in distorted visual encodings during prompt interactions. To address the above issues, we propose SAQ-SAM in this paper, which boosts PTQ of SAM with semantic alignment. Specifically, we propose Perceptual-Consistency Clipping, which exploits attention focus overlap as clipping metric, to significantly suppress outliers. Furthermore, we propose Prompt-Aware Reconstruction, which incorporates visual-prompt interactions by leveraging cross-attention responses in mask decoder, thus facilitating alignment in both distribution and semantics. To ensure the interaction efficiency, we also introduce a layer-skipping strategy for visual tokens. Extensive experiments are conducted on different segmentation tasks and SAMs of various sizes, and the results show that the proposed SAQ-SAM consistently outperforms baselines. For example, when quantizing SAM-B to 4-bit, our method achieves 11.7% higher mAP than the baseline in instance segmentation task.
Computational pathology and whole-slide image (WSI) analysis are pivotal in cancer diagnosis and prognosis. However, the ultra-high resolution of WSIs presents significant modeling challenges. Recent advancements in pathology foundation models have improved performance, yet most approaches rely on [CLS] token representation of tile ViT as slide-level inputs (16x16 pixels is refereed as patch and 224x224 pixels as tile). This discards critical spatial details from patch tokens, limiting downstream WSI analysis tasks. We find that leveraging all spatial patch tokens benefits WSI analysis but incurs nearly 200x higher storage and training costs (e.g., 196 tokens in ViT$_{224}$). To address this, we introduce vector quantized (VQ) distillation on patch feature, which efficiently compresses spatial patch tokens using discrete indices and a decoder. Our method reduces token dimensionality from 1024 to 16, achieving a 64x compression rate while preserving reconstruction fidelity. Furthermore, we employ a multi-scale VQ (MSVQ) strategy, which not only enhances VQ reconstruction performance but also serves as a Self-supervised Learning (SSL) supervision for a seamless slide-level pretraining objective. Built upon the quantized patch features and supervision targets of tile via MSVQ, we develop a progressive convolutional module and slide-level SSL to extract representations with rich spatial-information for downstream WSI tasks. Extensive evaluations on multiple datasets demonstrate the effectiveness of our approach, achieving state-of-the-art performance in WSI analysis. Code will be available soon.
In this work, we propose an extreme compression technique for Large Multimodal Models (LMMs). While previous studies have explored quantization as an efficient post-training compression method for Large Language Models (LLMs), low-bit compression for multimodal models remains under-explored. The redundant nature of inputs in multimodal models results in a highly sparse attention matrix. We theoretically and experimentally demonstrate that the attention matrix's sparsity bounds the compression error of the Query and Key weight matrices. Based on this, we introduce CASP, a model compression technique for LMMs. Our approach performs a data-aware low-rank decomposition on the Query and Key weight matrix, followed by quantization across all layers based on an optimal bit allocation process. CASP is compatible with any quantization technique and enhances state-of-the-art 2-bit quantization methods (AQLM and QuIP#) by an average of 21% on image- and video-language benchmarks.
Learnable Image Compression (LIC) has shown the potential to outperform standardized video codecs in RD efficiency, prompting the research for hardware-friendly implementations. Most existing LIC hardware implementations prioritize latency to RD-efficiency and through an extensive exploration of the hardware design space. We present a novel design paradigm where the burden of tuning the design for a specific hardware platform is shifted towards model dimensioning and without compromising on RD-efficiency. First, we design a framework for distilling a leaner student LIC model from a reference teacher: by tuning a single model hyperparameters, we can meet the constraints of different hardware platforms without a complex hardware design exploration. Second, we propose a hardware-friendly implementation of the Generalized Divisive Normalization - GDN activation that preserves RD efficiency even post parameter quantization. Third, we design a pipelined FPGA configuration which takes full advantage of available FPGA resources by leveraging parallel processing and optimizing resource allocation. Our experiments with a state of the art LIC model show that we outperform all existing FPGA implementations while performing very close to the original model.
Deep neural network (NN) with millions or billions of parameters can perform really well on unseen data, after being trained from a finite training set. Various prior theories have been developed to explain such excellent ability of NNs, but do not provide a meaningful bound on the test error. Some recent theories, based on PAC-Bayes and mutual information, are non-vacuous and hence show a great potential to explain the excellent performance of NNs. However, they often require a stringent assumption and extensive modification (e.g. compression, quantization) to the trained model of interest. Therefore, those prior theories provide a guarantee for the modified versions only. In this paper, we propose two novel bounds on the test error of a model. Our bounds uses the training set only and require no modification to the model. Those bounds are verified on a large class of modern NNs, pretrained by Pytorch on the ImageNet dataset, and are non-vacuous. To the best of our knowledge, these are the first non-vacuous bounds at this large scale, without any modification to the pretrained models.
Model merging enables efficient multi-task models by combining task-specific fine-tuned checkpoints. However, storing multiple task-specific checkpoints requires significant memory, limiting scalability and restricting model merging to larger models and diverse tasks. In this paper, we propose quantizing task vectors (i.e., the difference between pre-trained and fine-tuned checkpoints) instead of quantizing fine-tuned checkpoints. We observe that task vectors exhibit a narrow weight range, enabling low precision quantization (up to 4 bit) within existing task vector merging frameworks. To further mitigate quantization errors within ultra-low bit precision (e.g., 2 bit), we introduce Residual Task Vector Quantization, which decomposes the task vector into a base vector and offset component. We allocate bits based on quantization sensitivity, ensuring precision while minimizing error within a memory budget. Experiments on image classification and dense prediction show our method maintains or improves model merging performance while using only 8% of the memory required for full-precision checkpoints.
Large Vision and Language Models have exhibited remarkable human-like intelligence in tasks such as natural language comprehension, problem-solving, logical reasoning, and knowledge retrieval. However, training and serving these models require substantial computational resources, posing a significant barrier to their widespread application and further research. To mitigate this challenge, various model compression techniques have been developed to reduce computational requirements. Nevertheless, existing methods often employ uniform quantization configurations, failing to account for the varying difficulties across different layers in quantizing large neural network models. This paper tackles this issue by leveraging layer-sensitivity features, such as activation sensitivity and weight distribution Kurtosis, to identify layers that are challenging to quantize accurately and allocate additional memory budget. The proposed methods, named SensiBoost and KurtBoost, respectively, demonstrate notable improvement in quantization accuracy, achieving up to 9% lower perplexity with only a 2% increase in memory budget on LLama models compared to the baseline.