Vision-Language Models (VLMs) are powerful yet computationally intensive for widespread practical deployments. To address such challenge without costly re-training, post-training acceleration techniques like quantization and token reduction are extensively explored. However, current acceleration evaluations primarily target minimal overall performance degradation, overlooking a crucial question: does the accelerated model still give the same answers to the same questions as it did before acceleration? This is vital for stability-centered industrial applications where consistently correct answers for specific, known situations are paramount, such as in AI-based disease diagnosis. We systematically investigate this for accelerated VLMs, testing four leading models (LLaVA-1.5, LLaVA-Next, Qwen2-VL, Qwen2.5-VL) with eight acceleration methods on ten multi-modal benchmarks. Our findings are stark: despite minimal aggregate performance drops, accelerated models changed original answers up to 20% of the time. Critically, up to 6.5% of these changes converted correct answers to incorrect. Input perturbations magnified these inconsistencies, and the trend is confirmed by case studies with the medical VLM LLaVA-Med. This research reveals a significant oversight in VLM acceleration, stressing an urgent need for instance-level stability checks to ensure trustworthy real-world deployment.
Full 3D human pose reconstruction is a critical enabler for extended reality (XR) applications in future sixth generation (6G) networks, supporting immersive interactions in gaming, virtual meetings, and remote collaboration. However, achieving accurate pose reconstruction over wireless networks remains challenging due to channel impairments, bit errors, and quantization effects. Existing approaches often assume error-free transmission in indoor settings, limiting their applicability to real-world scenarios. To address these challenges, we propose a novel deep learning-based framework for human pose reconstruction over orthogonal frequency-division multiplexing (OFDM) systems. The framework introduces a two-stage deep learning receiver: the first stage jointly estimates the wireless channel and decodes OFDM symbols, and the second stage maps the received sensor signals to full 3D body poses. Simulation results demonstrate that the proposed neural receiver reduces bit error rate (BER), thus gaining a 5 dB gap at $10^{-4}$ BER, compared to the baseline method that employs separate signal detection steps, i.e., least squares channel estimation and linear minimum mean square error equalization. Additionally, our empirical findings show that 8-bit quantization is sufficient for accurate pose reconstruction, achieving a mean squared error of $5\times10^{-4}$ for reconstructed sensor signals, and reducing joint angular error by 37\% for the reconstructed human poses compared to the baseline.
With transformer-based models and the pretrain-finetune paradigm becoming mainstream, the high storage and deployment costs of individual finetuned models on multiple tasks pose critical challenges. Delta compression attempts to lower the costs by reducing the redundancy of delta parameters (i.e., the difference between the finetuned and pre-trained model weights). However, existing methods usually face problems including data accessibility and training requirements. To tackle this issue, we introduce Delta-DCT, the first data-free delta compression method inspired by classic JPEG image compression, leveraging the Discrete Cosine Transform (DCT). We first (a) group delta parameters within a layer into patches. Then we (b) assess the importance of each patch and allocate them with different quantization bit-widths. Afterwards, we (c) convert these patches to the DCT domain and conduct quantization to each patch based on the allocated bit-width. The proposed Delta-DCT does not require any training or data calibration, while achieving performance comparable to or even surpassing original finetuned models under 1-bit equivalent delta compression ratios on different kinds of models including: (1) recently-released LLMs of different sizes from 7B to 13B, (2) relatively smaller language models including RoBERTa and T5 models, (3) variants of vision transformer models, and (4) multi-modal BEiT-3 models.
Video-to-audio generation is essential for synthesizing realistic audio tracks that synchronize effectively with silent videos. Following the perspective of extracting essential signals from videos that can precisely control the mature text-to-audio generative diffusion models, this paper presents how to balance the representation of mel-spectrograms in terms of completeness and complexity through a new approach called Mel Quantization-Continuum Decomposition (Mel-QCD). We decompose the mel-spectrogram into three distinct types of signals, employing quantization or continuity to them, we can effectively predict them from video by a devised video-to-all (V2X) predictor. Then, the predicted signals are recomposed and fed into a ControlNet, along with a textual inversion design, to control the audio generation process. Our proposed Mel-QCD method demonstrates state-of-the-art performance across eight metrics, evaluating dimensions such as quality, synchronization, and semantic consistency. Our codes and demos will be released at \href{Website}{https://wjc2830.github.io/MelQCD/}.
Maritime Multi-Scene Recognition is crucial for enhancing the capabilities of intelligent marine robotics, particularly in applications such as marine conservation, environmental monitoring, and disaster response. However, this task presents significant challenges due to environmental interference, where marine conditions degrade image quality, and the complexity of maritime scenes, which requires deeper reasoning for accurate recognition. Pure vision models alone are insufficient to address these issues. To overcome these limitations, we propose a novel multimodal Artificial Intelligence (AI) framework that integrates image data, textual descriptions and classification vectors generated by a Multimodal Large Language Model (MLLM), to provide richer semantic understanding and improve recognition accuracy. Our framework employs an efficient multimodal fusion mechanism to further enhance model robustness and adaptability in complex maritime environments. Experimental results show that our model achieves 98$\%$ accuracy, surpassing previous SOTA models by 3.5$\%$. To optimize deployment on resource-constrained platforms, we adopt activation-aware weight quantization (AWQ) as a lightweight technique, reducing the model size to 68.75MB with only a 0.5$\%$ accuracy drop while significantly lowering computational overhead. This work provides a high-performance solution for real-time maritime scene recognition, enabling Autonomous Surface Vehicles (ASVs) to support environmental monitoring and disaster response in resource-limited settings.
Learning to perform accurate and rich simulations of human driving behaviors from data for autonomous vehicle testing remains challenging due to human driving styles' high diversity and variance. We address this challenge by proposing a novel approach that leverages contrastive learning to extract a dictionary of driving styles from pre-existing human driving data. We discretize these styles with quantization, and the styles are used to learn a conditional diffusion policy for simulating human drivers. Our empirical evaluation confirms that the behaviors generated by our approach are both safer and more human-like than those of the machine-learning-based baseline methods. We believe this has the potential to enable higher realism and more effective techniques for evaluating and improving the performance of autonomous vehicles.
We present a novel approach to selective model quantization that transcends the limitations of architecture-specific and size-dependent compression methods for Large Language Models (LLMs) using Entropy-Weighted Quantization (EWQ). By analyzing the entropy distribution across transformer blocks, EWQ determines which blocks can be safely quantized without causing significant performance degradation, independent of model architecture or size. Our method outperforms uniform quantization approaches, maintaining Massive Multitask Language Understanding (MMLU) accuracy scores within 0.5% of unquantized models while reducing memory usage by up to 18%. We demonstrate the effectiveness of EWQ across multiple architectures-from 1.6B to 70B parameters-showcasing consistent improvements in the quality-compression trade-off regardless of model scale or architectural design. A surprising finding of EWQ is its ability to reduce perplexity compared to unquantized models, suggesting the presence of beneficial regularization through selective precision reduction. This improvement holds across different model families, indicating a fundamental relationship between layer-level entropy and optimal precision requirements. Additionally, we introduce FastEWQ, a rapid method for entropy distribution analysis that eliminates the need for loading model weights. This technique leverages universal characteristics of entropy distribution that persist across various architectures and scales, enabling near-instantaneous quantization decisions while maintaining 80% classification accuracy with full entropy analysis. Our results demonstrate that effective quantization strategies can be developed independently of specific architectural choices or model sizes, opening new possibilities for efficient LLM deployment.
The optimal bit-width for achieving the best trade-off between quantized model size and accuracy has been a subject of ongoing debate. While some advocate for 4-bit quantization, others propose that 1.58-bit offers superior results. However, the lack of a cohesive framework for different bits has left such conclusions relatively tenuous. We present ParetoQ, the first unified framework that facilitates rigorous comparisons across 1-bit, 1.58-bit, 2-bit, 3-bit, and 4-bit quantization settings. Our findings reveal a notable learning transition between 2 and 3 bits: For 3-bits and above, the fine-tuned models stay close to their original pre-trained distributions, whereas for learning 2-bit networks or below, the representations change drastically. By optimizing training schemes and refining quantization functions, ParetoQ surpasses all previous methods tailored to specific bit widths. Remarkably, our ParetoQ ternary 600M-parameter model even outperforms the previous SoTA ternary 3B-parameter model in accuracy, using only one-fifth of the parameters. Extensive experimentation shows that ternary, 2-bit, and 3-bit quantization maintains comparable performance in the size-accuracy trade-off and generally exceeds 4-bit and binary quantization. Considering hardware constraints, 2-bit quantization offers promising potential for memory reduction and speedup.
For consumer usage of locally deployed LLMs, the GGUF format and k_quantization are invaluable tools for maintaining the performance of the original model while reducing it to sizes deployable with consumer-grade hardware. The number of bits dedicated to each weight from the original model is reduced based on how important they are thought to be during model inference. This importance is arrived at through the application of an 'importance matrix'-a relatively small text document meant to be representative of the LLM's standard use-cases. In the vast majority of quants available online, this document is primarily written in English. It was therefore an open question whether performance on English language tasks was preserved through the sacrifice of multilingual performance and whether it can be preserved with alternate importance matrices. This article investigates these hypotheses by quantizing Llama3.3 70B on importance matrices written in three languages (English, Norwegian, and Malayalam) and evaluating them on the MixEval dataset in both English and Norwegian. All experiments related to k_quantization yielded non-significant results (In all cases p > 0.237) indicating that current quantization practices do not disproportionately harm multilingual performance.
Model binarization has made significant progress in enabling real-time and energy-efficient computation for convolutional neural networks (CNN), offering a potential solution to the deployment challenges faced by Vision Transformers (ViTs) on edge devices. However, due to the structural differences between CNN and Transformer architectures, simply applying binary CNN strategies to the ViT models will lead to a significant performance drop. To tackle this challenge, we propose BHViT, a binarization-friendly hybrid ViT architecture and its full binarization model with the guidance of three important observations. Initially, BHViT utilizes the local information interaction and hierarchical feature aggregation technique from coarse to fine levels to address redundant computations stemming from excessive tokens. Then, a novel module based on shift operations is proposed to enhance the performance of the binary Multilayer Perceptron (MLP) module without significantly increasing computational overhead. In addition, an innovative attention matrix binarization method based on quantization decomposition is proposed to evaluate the token's importance in the binarized attention matrix. Finally, we propose a regularization loss to address the inadequate optimization caused by the incompatibility between the weight oscillation in the binary layers and the Adam Optimizer. Extensive experimental results demonstrate that our proposed algorithm achieves SOTA performance among binary ViT methods.
We investigate a class of non-Markovian processes that hold particular relevance in the realm of mathematical finance. This family encompasses path-dependent volatility models, including those pioneered by [Platen and Rendek, 2018] and, more recently, by [Guyon and Lekeufack, 2023], as well as an extension of the framework proposed by [Blanc et al., 2017]. Our study unfolds in two principal phases. In the first phase, we introduce a functional quantization scheme based on an extended version of the Lamperti transformation that we propose to handle the presence of a memory term incorporated into the diffusion coefficient. For scenarios involving a Brownian integral in the diffusion term, we propose alternative numerical schemes that leverage the power of marginal recursive quantization. In the second phase, we study the problem of existence and uniqueness of a strong solution for the SDEs related to the examples that motivate our study, in order to provide a theoretical basis to correctly apply the proposed numerical schemes.
Quantization and cache mechanisms are typically applied individually for efficient Diffusion Transformers (DiTs), each demonstrating notable potential for acceleration. However, the promoting effect of combining the two mechanisms on efficient generation remains under-explored. Through empirical investigation, we find that the combination of quantization and cache mechanisms for DiT is not straightforward, and two key challenges lead to severe catastrophic performance degradation: (i) the sample efficacy of calibration datasets in post-training quantization (PTQ) is significantly eliminated by cache operation; (ii) the combination of the above mechanisms introduces more severe exposure bias within sampling distribution, resulting in amplified error accumulation in the image generation process. In this work, we take advantage of these two acceleration mechanisms and propose a hybrid acceleration method by tackling the above challenges, aiming to further improve the efficiency of DiTs while maintaining excellent generation capability. Concretely, a temporal-aware parallel clustering (TAP) is designed to dynamically improve the sample selection efficacy for the calibration within PTQ for different diffusion steps. A variance compensation (VC) strategy is derived to correct the sampling distribution. It mitigates exposure bias through an adaptive correction factor generation. Extensive experiments have shown that our method has accelerated DiTs by 12.7x while preserving competitive generation capability. The code will be available at https://github.com/xinding-sys/Quant-Cache.
Neural audio signal codecs have attracted significant attention in recent years. In essence, the impressive low bitrate achieved by such encoders is enabled by learning an abstract representation that captures the properties of encoded signals, e.g., speech. In this work, we investigate the relation between the latent representation of the input signal learned by a neural codec and the quality of speech signals. To do so, we introduce Latent-representation-to-Quantization error Ratio (LQR) measures, which quantify the distance from the idealized neural codec's speech signal model for a given speech signal. We compare the proposed metrics to intrusive measures as well as data-driven supervised methods using two subjective speech quality datasets. This analysis shows that the proposed LQR correlates strongly (up to 0.9 Pearson's correlation) with the subjective quality of speech. Despite being a non-intrusive metric, this yields a competitive performance with, or even better than, other pre-trained and intrusive measures. These results show that LQR is a promising basis for more sophisticated speech quality measures.
We explore the innovative use of MLP-Mixer models for real-time jet tagging and establish their feasibility on resource-constrained hardware like FPGAs. MLP-Mixers excel in processing sequences of jet constituents, achieving state-of-the-art performance on datasets mimicking Large Hadron Collider conditions. By using advanced optimization techniques such as High-Granularity Quantization and Distributed Arithmetic, we achieve unprecedented efficiency. These models match or surpass the accuracy of previous architectures, reduce hardware resource usage by up to 97%, double the throughput, and half the latency. Additionally, non-permutation-invariant architectures enable smart feature prioritization and efficient FPGA deployment, setting a new benchmark for machine learning in real-time data processing at particle colliders.
Sign language recognition (SLR) for low-resource languages like Bangla suffers from signer variability, viewpoint variations, and limited annotated datasets. In this paper, we present BdSLW401, a large-scale, multi-view, word-level Bangla Sign Language (BdSL) dataset with 401 signs and 102,176 video samples from 18 signers in front and lateral views. To improve transformer-based SLR, we introduce Relative Quantization Encoding (RQE), a structured embedding approach anchoring landmarks to physiological reference points and quantize motion trajectories. RQE improves attention allocation by decreasing spatial variability, resulting in 44.3% WER reduction in WLASL100, 21.0% in SignBD-200, and significant gains in BdSLW60 and SignBD-90. However, fixed quantization becomes insufficient on large-scale datasets (e.g., WLASL2000), indicating the need for adaptive encoding strategies. Further, RQE-SF, an extended variant that stabilizes shoulder landmarks, achieves improvements in pose consistency at the cost of small trade-offs in lateral view recognition. The attention graphs prove that RQE improves model interpretability by focusing on the major articulatory features (fingers, wrists) and the more distinctive frames instead of global pose changes. Introducing BdSLW401 and demonstrating the effectiveness of RQE-enhanced structured embeddings, this work advances transformer-based SLR for low-resource languages and sets a benchmark for future research in this area.
The Segment Anything Model (SAM) has demonstrated strong versatility across various visual tasks. However, its large storage requirements and high computational cost pose challenges for practical deployment. Post-training quantization (PTQ) has emerged as an effective strategy for efficient deployment, but we identify two key challenges in SAM that hinder the effectiveness of existing PTQ methods: the heavy-tailed and skewed distribution of post-GELU activations, and significant inter-channel variation in linear projection activations. To address these challenges, we propose AHCPTQ, an accurate and hardware-efficient PTQ method for SAM. AHCPTQ introduces hardware-compatible Hybrid Log-Uniform Quantization (HLUQ) to manage post-GELU activations, employing log2 quantization for dense small values and uniform quantization for sparse large values to enhance quantization resolution. Additionally, AHCPTQ incorporates Channel-Aware Grouping (CAG) to mitigate inter-channel variation by progressively clustering activation channels with similar distributions, enabling them to share quantization parameters and improving hardware efficiency. The combination of HLUQ and CAG not only enhances quantization effectiveness but also ensures compatibility with efficient hardware execution. For instance, under the W4A4 configuration on the SAM-L model, AHCPTQ achieves 36.6% mAP on instance segmentation with the DINO detector, while achieving a 7.89x speedup and 8.64x energy efficiency over its floating-point counterpart in FPGA implementation.
With a recent trend of using Large Language Models (LLMs) for different applications within smart cities, there is a need for pushing these models toward the edge of network while still preserving their performance. Edge Computing (EC) as a physically closer computing resource to the end users can help to reduce the communication delay for serving end users' tasks for LLM-dependent services. However, EC servers have limited capacity in terms of communication, computation, and storage capacity. This paper introduces DILEMMA, a novel framework addressing the challenges of deploying LLMs in EC systems by jointly optimizing layer placement and layer quantization in EC systems. DILEMMA formulates an Integer Linear Programming problem to minimize total inference delay while ensuring acceptable LLM performance levels, leveraging layer-wise quantization and knowledge distillation for LLM performance control. Experimental evaluations on OPT-350 model using the SQuAD dataset demonstrate that DILEMMA achieves a quantization ratio of up to 12.75% while preserving model loss, highlighting its effectiveness in resource-constrained environments.
We introduce a novel class of regularization functions, called Cauchy-Schwarz (CS) regularizers, which can be designed to induce a wide range of properties in solution vectors of optimization problems. To demonstrate the versatility of CS regularizers, we derive regularization functions that promote discrete-valued vectors, eigenvectors of a given matrix, and orthogonal matrices. The resulting CS regularizers are simple, differentiable, and can be free of spurious stationary points, making them suitable for gradient-based solvers and large-scale optimization problems. In addition, CS regularizers automatically adapt to the appropriate scale, which is, for example, beneficial when discretizing the weights of neural networks. To demonstrate the efficacy of CS regularizers, we provide results for solving underdetermined systems of linear equations and weight quantization in neural networks. Furthermore, we discuss specializations, variations, and generalizations, which lead to an even broader class of new and possibly more powerful regularizers.
One of the challenges of quantizing a large language model (LLM) is the presence of outliers. Outliers often make uniform quantization schemes less effective, particularly in extreme cases such as 4-bit quantization. We introduce KurTail, a new post-training quantization (PTQ) scheme that leverages Kurtosis-based rotation to mitigate outliers in the activations of LLMs. Our method optimizes Kurtosis as a measure of tailedness. This approach enables the quantization of weights, activations, and the KV cache in 4 bits. We utilize layer-wise optimization, ensuring memory efficiency. KurTail outperforms existing quantization methods, offering a 13.3\% boost in MMLU accuracy and a 15.5\% drop in Wiki perplexity compared to QuaRot. It also outperforms SpinQuant with a 2.6\% MMLU gain and reduces perplexity by 2.9\%, all while reducing the training cost. For comparison, learning the rotation using SpinQuant for Llama3-70B requires at least four NVIDIA H100 80GB GPUs, whereas our method requires only a single GPU, making it a more accessible solution for consumer GPU.
Upcycled Mixture-of-Experts (MoE) models have shown great potential in various tasks by converting the original Feed-Forward Network (FFN) layers in pre-trained dense models into MoE layers. However, these models still suffer from significant parameter inefficiency due to the introduction of multiple experts. In this work, we propose a novel DeRS (Decompose, Replace, and Synthesis) paradigm to overcome this shortcoming, which is motivated by our observations about the unique redundancy mechanisms of upcycled MoE experts. Specifically, DeRS decomposes the experts into one expert-shared base weight and multiple expert-specific delta weights, and subsequently represents these delta weights in lightweight forms. Our proposed DeRS paradigm can be applied to enhance parameter efficiency in two different scenarios, including: 1) DeRS Compression for inference stage, using sparsification or quantization to compress vanilla upcycled MoE models; and 2) DeRS Upcycling for training stage, employing lightweight sparse or low-rank matrixes to efficiently upcycle dense models into MoE models. Extensive experiments across three different tasks show that the proposed methods can achieve extreme parameter efficiency while maintaining the performance for both training and compression of upcycled MoE models.