The impressive capabilities of Large Language Models (LLMs) come at the cost of substantial computational resources during deployment. While KV Cache can significantly reduce recomputation during inference, it also introduces additional memory overhead. KV Cache quantization presents a promising solution, striking a good balance between memory usage and accuracy. Previous research has shown that the Keys are distributed by channel, while the Values are distributed by token. Consequently, the common practice is to apply channel-wise quantization to the Keys and token-wise quantization to the Values. However, our further investigation reveals that a small subset of unusual tokens exhibit unique characteristics that deviate from this pattern, which can substantially impact quantization accuracy. To address this, we develop a simple yet effective method to identify these tokens accurately during the decoding process and exclude them from quantization as outlier tokens, significantly improving overall accuracy. Extensive experiments show that our method achieves significant accuracy improvements under 2-bit quantization and can deliver a 6.4 times reduction in memory usage and a 2.3 times increase in throughput.
Ever-growing scale of large language models (LLMs) is pushing for improved efficiency, favoring fully quantized training (FQT) over BF16. While FQT accelerates training, it faces consistency challenges and requires searching over an exponential number of cases, each needing over 200B tokens to ensure stability. Pseudo-quantization training (PQT) addresses the issues of FQT, although it is not well-studied. We explore the practical implications of PQT in detail and propose a noise distribution $R$ that is floating-point (FP)-friendly, with ideal properties including stochastic precision annealing. As a result, the proposed method serves as an effective theoretical foundation for low-precision FP parameters through PQT, utilizing efficient fake quantization via an addition and subsequent FP casting. We demonstrate that Gaussian weight sampling is (1) scalable: supports low-precision FP parameters down to FP6 and high-precision noise up to 9-bit with BF16 operator. The proposed method is (2) efficient: incurring computational overhead as low as 1.40\% on the A100 GPU in terms of Llama2 training tokens per second, and requiring 2 bytes per parameter in GPU memory. We demonstrate that PQT with Gaussian weight sampling is (3) stable: closely following or even surpassing performance of the BF16 baseline while pre-training GPT2 and Llama2 models with up to 1B parameters and 300B tokens.
Binary quantization approaches, which replace weight matrices with binary matrices and substitute costly multiplications with cheaper additions, offer a computationally efficient approach to address the increasing computational and storage requirements of Large Language Models (LLMs). However, the severe quantization constraint ($\pm1$) can lead to significant accuracy degradation. In this paper, we propose Double Binary Factorization (DBF), a novel method that factorizes dense weight matrices into products of two binary (sign) matrices, each accompanied by scaling vectors. DBF preserves the efficiency advantages of binary representations while achieving compression rates that are competitive with or superior to state-of-the-art methods. Specifically, in a 1-bit per weight range, DBF is better than existing binarization approaches. In a 2-bit per weight range, DBF is competitive with the best quantization methods like QuIP\# and QTIP. Unlike most existing compression techniques, which offer limited compression level choices, DBF allows fine-grained control over compression ratios by adjusting the factorization's intermediate dimension. Based on this advantage, we further introduce an algorithm for estimating non-uniform layer-wise compression ratios for DBF, based on previously developed channel pruning criteria. Code available at: https://github.com/usamec/double_binary
We propose the first unified adversarial attack benchmark for Genomic Foundation Models (GFMs), named GenoArmory. Unlike existing GFM benchmarks, GenoArmory offers the first comprehensive evaluation framework to systematically assess the vulnerability of GFMs to adversarial attacks. Methodologically, we evaluate the adversarial robustness of five state-of-the-art GFMs using four widely adopted attack algorithms and three defense strategies. Importantly, our benchmark provides an accessible and comprehensive framework to analyze GFM vulnerabilities with respect to model architecture, quantization schemes, and training datasets. Additionally, we introduce GenoAdv, a new adversarial sample dataset designed to improve GFM safety. Empirically, classification models exhibit greater robustness to adversarial perturbations compared to generative models, highlighting the impact of task type on model vulnerability. Moreover, adversarial attacks frequently target biologically significant genomic regions, suggesting that these models effectively capture meaningful sequence features.
Video diffusion models (DMs) have enabled high-quality video synthesis. Yet, their substantial computational and memory demands pose serious challenges to real-world deployment, even on high-end GPUs. As a commonly adopted solution, quantization has proven notable success in reducing cost for image DMs, while its direct application to video DMs remains ineffective. In this paper, we present QVGen, a novel quantization-aware training (QAT) framework tailored for high-performance and inference-efficient video DMs under extremely low-bit quantization (e.g., 4-bit or below). We begin with a theoretical analysis demonstrating that reducing the gradient norm is essential to facilitate convergence for QAT. To this end, we introduce auxiliary modules ($\Phi$) to mitigate large quantization errors, leading to significantly enhanced convergence. To eliminate the inference overhead of $\Phi$, we propose a rank-decay strategy that progressively eliminates $\Phi$. Specifically, we repeatedly employ singular value decomposition (SVD) and a proposed rank-based regularization $\mathbf{\gamma}$ to identify and decay low-contributing components. This strategy retains performance while zeroing out inference overhead. Extensive experiments across $4$ state-of-the-art (SOTA) video DMs, with parameter sizes ranging from $1.3$B $\sim14$B, show that QVGen is the first to reach full-precision comparable quality under 4-bit settings. Moreover, it significantly outperforms existing methods. For instance, our 3-bit CogVideoX-2B achieves improvements of $+25.28$ in Dynamic Degree and $+8.43$ in Scene Consistency on VBench.
This work aims at a challenging task: human action-reaction synthesis, i.e., generating human reactions based on the action sequence of the other as conditions. Currently, autoregressive modeling approaches have achieved remarkable performance in motion generation tasks, e.g. text-to-motion. However, vector quantization (VQ) accompanying autoregressive generation has inherent disadvantages, including loss of quantization information, low codebook utilization, etc. Moreover, unlike text-to-motion, which focuses solely on the movement of body joints, human action-reaction synthesis also encompasses fine-grained hand movements. In this work, we propose MARRS, a novel framework designed to generate coordinated and fine-grained reaction motions in continuous representations. Initially, we present the Unit-distinguished Motion Variational AutoEncoder (UD-VAE), which segments the entire body into distinct body and hand units, encoding them independently. Subsequently, we propose Action-Conditioned Fusion (ACF), which involves randomly masking a subset of reactive tokens and extracting specific information about the body and hands from the active tokens. Furthermore, we introduce Adaptive Unit Modulation (AUM) to facilitate interaction between body and hand units by using the information from one unit to adaptively modulate the other. Finally, for the diffusion model, we employ a compact MLP as a noise predictor for each distinct body unit and incorporate the diffusion loss to model the probability distribution of each token. Quantitative and qualitative results demonstrate that our method achieves superior performance. The code will be released upon acceptance.
We propose a probabilistic framework for dynamic quantization of neural networks that allows for a computationally efficient input-adaptive rescaling of the quantization parameters. Our framework applies a probabilistic model to the network's pre-activations through a lightweight surrogate, enabling the adaptive adjustment of the quantization parameters on a per-input basis without significant memory overhead. We validate our approach on a set of popular computer vision tasks and models, observing only a negligible loss in performance. Our method strikes the best performance and computational overhead tradeoff compared to standard quantization strategies.
Efficient scene representations are essential for many real-world applications, especially those involving spatial measurement. Although current NeRF-based methods have achieved impressive results in reconstructing building-scale scenes, they still suffer from slow training and inference speeds due to time-consuming stochastic sampling. Recently, 3D Gaussian Splatting (3DGS) has demonstrated excellent performance with its high-quality rendering and real-time speed, especially for objects and small-scale scenes. However, in outdoor scenes, its point-based explicit representation lacks an effective adjustment mechanism, and the millions of Gaussian points required often lead to memory constraints during training. To address these challenges, we propose EA-3DGS, a high-quality real-time rendering method designed for outdoor scenes. First, we introduce a mesh structure to regulate the initialization of Gaussian components by leveraging an adaptive tetrahedral mesh that partitions the grid and initializes Gaussian components on each face, effectively capturing geometric structures in low-texture regions. Second, we propose an efficient Gaussian pruning strategy that evaluates each 3D Gaussian's contribution to the view and prunes accordingly. To retain geometry-critical Gaussian points, we also present a structure-aware densification strategy that densifies Gaussian points in low-curvature regions. Additionally, we employ vector quantization for parameter quantization of Gaussian components, significantly reducing disk space requirements with only a minimal impact on rendering quality. Extensive experiments on 13 scenes, including eight from four public datasets (MatrixCity-Aerial, Mill-19, Tanks \& Temples, WHU) and five self-collected scenes acquired through UAV photogrammetry measurement from SCUT-CA and plateau regions, further demonstrate the superiority of our method.
Large Language Models (LLMs) have achieved remarkable success but face significant computational and memory challenges, particularly due to their extensive output vocabularies. The final linear projection layer, mapping hidden states to vocabulary-sized logits, often constitutes a substantial portion of the model's parameters and computational cost during inference. Existing methods like adaptive softmax or hierarchical softmax introduce structural complexities. In this paper, we propose VQ-Logits, a novel approach that leverages Vector Quantization (VQ) to drastically reduce the parameter count and computational load of the LLM output layer. VQ-Logits replaces the large V * dmodel output embedding matrix with a small, shared codebook of K embedding vectors (K << V ). Each token in the vocabulary is mapped to one of these K codebook vectors. The LLM predicts logits over this compact codebook, which are then efficiently "scattered" to the full vocabulary space using the learned or preassigned mapping. We demonstrate through extensive experiments on standard language modeling benchmarks (e.g., WikiText-103, C4) that VQ-Logits can achieve up to 99% parameter reduction in the output layer and 6x speedup in logit computation, with only a marginal 4% increase in perplexity compared to full softmax baselines. We further provide detailed ablation studies on codebook size, initialization, and learning strategies, showcasing the robustness and effectiveness of our approach.
Analog in-memory computing (AIMC) is a promising compute paradigm to improve speed and power efficiency of neural network inference beyond the limits of conventional von Neumann-based architectures. However, AIMC introduces fundamental challenges such as noisy computations and strict constraints on input and output quantization. Because of these constraints and imprecisions, off-the-shelf LLMs are not able to achieve 4-bit-level performance when deployed on AIMC-based hardware. While researchers previously investigated recovering this accuracy gap on small, mostly vision-based models, a generic method applicable to LLMs pre-trained on trillions of tokens does not yet exist. In this work, we introduce a general and scalable method to robustly adapt LLMs for execution on noisy, low-precision analog hardware. Our approach enables state-of-the-art models $\unicode{x2013}$ including Phi-3-mini-4k-instruct and Llama-3.2-1B-Instruct $\unicode{x2013}$ to retain performance comparable to 4-bit weight, 8-bit activation baselines, despite the presence of analog noise and quantization constraints. Additionally, we show that as a byproduct of our training methodology, analog foundation models can be quantized for inference on low-precision digital hardware. Finally, we show that our models also benefit from test-time compute scaling, showing better scaling behavior than models trained with 4-bit weight and 8-bit static input quantization. Our work bridges the gap between high-capacity LLMs and efficient analog hardware, offering a path toward energy-efficient foundation models. Code is available at https://github.com/IBM/analog-foundation-models.
Unsupervised domain adaptation (UDA) for time series data remains a critical challenge in deep learning, with traditional pseudo-labeling strategies failing to capture temporal patterns and channel-wise shifts between domains, producing sub-optimal pseudo-labels. As such, we introduce TransPL, a novel approach that addresses these limitations by modeling the joint distribution $P(\mathbf{X}, y)$ of the source domain through code transition matrices, where the codes are derived from vector quantization (VQ) of time series patches. Our method constructs class- and channel-wise code transition matrices from the source domain and employs Bayes' rule for target domain adaptation, generating pseudo-labels based on channel-wise weighted class-conditional likelihoods. TransPL offers three key advantages: explicit modeling of temporal transitions and channel-wise shifts between different domains, versatility towards different UDA scenarios (e.g., weakly-supervised UDA), and explainable pseudo-label generation. We validate TransPL's effectiveness through extensive analysis on four time series UDA benchmarks and confirm that it consistently outperforms state-of-the-art pseudo-labeling methods by a strong margin (6.1% accuracy improvement, 4.9% F1 improvement), while providing interpretable insights into the domain adaptation process through its learned code transition matrices.
Graph Neural Networks (GNNs) have become essential for handling large-scale graph applications. However, the computational demands of GNNs necessitate the development of efficient methods to accelerate inference. Mixed precision quantization emerges as a promising solution to enhance the efficiency of GNN architectures without compromising prediction performance. Compared to conventional deep learning architectures, GNN layers contain a wider set of components that can be quantized, including message passing functions, aggregation functions, update functions, the inputs, learnable parameters, and outputs of these functions. In this paper, we introduce a theorem for efficient quantized message passing to aggregate integer messages. It guarantees numerical equality of the aggregated messages using integer values with respect to those obtained with full (FP32) precision. Based on this theorem, we introduce the Mixed Precision Quantization for GNN (MixQ-GNN) framework, which flexibly selects effective integer bit-widths for all components within GNN layers. Our approach systematically navigates the wide set of possible bit-width combinations, addressing the challenge of optimizing efficiency while aiming at maintaining comparable prediction performance. MixQ-GNN integrates with existing GNN quantization methods, utilizing their graph structure advantages to achieve higher prediction performance. On average, MixQ-GNN achieved reductions in bit operations of 5.5x for node classification and 5.1x for graph classification compared to architectures represented in FP32 precision.
Network quantization has proven to be a powerful approach to reduce the memory and computational demands of deep learning models for deployment on resource-constrained devices. However, traditional quantization methods often rely on access to training data, which is impractical in many real-world scenarios due to privacy, security, or regulatory constraints. Zero-shot Quantization (ZSQ) emerges as a promising solution, achieving quantization without requiring any real data. In this paper, we provide a comprehensive overview of ZSQ methods and their recent advancements. First, we provide a formal definition of the ZSQ problem and highlight the key challenges. Then, we categorize the existing ZSQ methods into classes based on data generation strategies, and analyze their motivations, core ideas, and key takeaways. Lastly, we suggest future research directions to address the remaining limitations and advance the field of ZSQ. To the best of our knowledge, this paper is the first in-depth survey on ZSQ.
Large language models (LLMs) have transformed natural-language processing, yet their scale makes real-world deployment costly. Post-training quantization reduces memory and computation but often degrades accuracy, while quantization-aware training can recover performance at the cost of extra training. Pushing quantization to the ternary (2-bit) regime yields even larger savings but is notoriously unstable. Building on recent work showing that a bias-free, RMS-normalized Transformer with straight-through estimation can reach 1.58-bit precision, we demonstrate that simply inserting RMS normalization before every linear projection and applying a gradual, layer-wise quantization schedule stably fine-tunes full-precision checkpoints into ternary LLMs. Our approach matches or surpasses more elaborate knowledge-distillation pipelines on standard language-modeling benchmarks without adding model complexity. These results indicate that careful normalization alone can close much of the accuracy gap between ternary and full-precision LLMs, making ultra-low-bit inference practical.
This paper deals with the design of non-coherent jamming strategies capable of ensuring spectral compatibility with friendly radio frequency (RF) emitters. The goal is achieved via a cognitive approach, which, after recognizing the presence of friendly RF systems within the bandwidth of interest (perception), synthesizes a jamming waveform (action) with spectral notches, that allows to interfere exclusively with opposite emissions. Two methods are proposed for the synthesis of the jamming signal. The former leverages optimization techniques for quadratically constrained quadratic problems (QCQP) where each constraint embeds the interference level tolerable by a specific friendly RF system. The latter is a very computationally efficient approach based on simple projections, allowing a control over the spectral notch positions and widths. At the analysis stage, the performance of the devised jamming techniques is firstly numerically analyzed in terms of spectral occupancy and autocorrelation characteristics. The impact of the quantization process involved in the digital-to-analog conversion (DAC) of the jamming waveforms is also examined, with a particular focus on the spectral shaping impairments resulting from reduced DAC resolution. Finally, waveform transmission and reception is experimentally assessed with software defined radio (SDR) devices.
Almost all existing hierarchical federated learning (FL) models are limited to two aggregation layers, restricting scalability and flexibility in complex, large-scale networks. In this work, we propose a Multi-Layer Hierarchical Federated Learning framework (QMLHFL), which appears to be the first study that generalizes hierarchical FL to arbitrary numbers of layers and network architectures through nested aggregation, while employing a layer-specific quantization scheme to meet communication constraints. We develop a comprehensive convergence analysis for QMLHFL and derive a general convergence condition and rate that reveal the effects of key factors, including quantization parameters, hierarchical architecture, and intra-layer iteration counts. Furthermore, we determine the optimal number of intra-layer iterations to maximize the convergence rate while meeting a deadline constraint that accounts for both communication and computation times. Our results show that QMLHFL consistently achieves high learning accuracy, even under high data heterogeneity, and delivers notably improved performance when optimized, compared to using randomly selected values.
Large language models have significantly advanced natural language processing, yet their heavy resource demands pose severe challenges regarding hardware accessibility and energy consumption. This paper presents a focused and high-level review of post-training quantization (PTQ) techniques designed to optimize the inference efficiency of LLMs by the end-user, including details on various quantization schemes, granularities, and trade-offs. The aim is to provide a balanced overview between the theory and applications of post-training quantization.
Large language models (LLMs) demand extensive memory capacity during both fine-tuning and inference. To enable memory-efficient fine-tuning, existing methods apply block-wise quantization techniques, such as NF4 and AF4, to the network weights. We show that these quantization techniques incur suboptimal quantization errors. Therefore, as a first novelty, we propose an optimization approach for block-wise quantization. Using this method, we design a family of quantizers named 4-bit block-wise optimal float (BOF4), which consistently reduces the quantization error compared to both baseline methods. We provide both a theoretical and a data-driven solution for the optimization process and prove their practical equivalence. Secondly, we propose a modification to the employed normalization method based on the signed absolute block maximum (BOF4-S), enabling further reduction of the quantization error and empirically achieving less degradation in language modeling performance. Thirdly, we explore additional variations of block-wise quantization methods applied to LLMs through an experimental study on the importance of accurately representing zero and large-amplitude weights on the one hand, and optimization towards various error metrics on the other hand. Lastly, we introduce a mixed-precision quantization strategy dubbed outlier-preserving quantization (OPQ) to address the distributional mismatch induced by outlier weights in block-wise quantization. By storing outlier weights in 16-bit precision (OPQ) while applying BOF4-S, we achieve top performance among 4-bit block-wise quantization techniques w.r.t. perplexity.
The common assumption in on-device AI is that GPUs, with their superior parallel processing, always provide the best performance for large language model (LLM) inference. In this work, we challenge this notion by empirically demonstrating that, under certain conditions, CPUs can outperform GPUs for LLM inference on mobile devices. Using a 1-billion-parameter LLM deployed via llama.cpp on the iPhone 15 Pro, we show that a CPU-only configuration (two threads, F16 precision) achieves 17 tokens per second, surpassing the 12.8 tokens per second obtained with GPU acceleration. We analyze the architectural factors driving this counterintuitive result, revealing that GPU memory transfer overhead and CPU thread optimization play a critical role. Furthermore, we explore the impact of thread oversubscription, quantization strategies, and hardware constraints, providing new insights into efficient on-device AI execution. Our findings challenge conventional GPU-first thinking, highlighting the untapped potential of optimized CPU inference and paving the way for smarter deployment strategies in mobile AI. However, fully explaining the observed CPU advantage remains difficult due to limited access to low-level profiling tools on iOS.
Preserving data confidentiality during the fine-tuning of open-source Large Language Models (LLMs) is crucial for sensitive applications. This work introduces an interactive protocol adapting the Low-Rank Adaptation (LoRA) technique for private fine-tuning. Homomorphic Encryption (HE) protects the confidentiality of training data and gradients handled by remote worker nodes performing the bulk of computations involving the base model weights. The data owner orchestrates training, requiring minimal local computing power and memory, thus alleviating the need for expensive client-side GPUs. We demonstrate feasibility by fine-tuning a Llama-3.2-1B model, presenting convergence results using HE-compatible quantization and performance benchmarks for HE computations on GPU hardware. This approach enables applications such as confidential knowledge base question answering, private codebase fine-tuning for AI code assistants, AI agents for drafting emails based on a company's email archive, and adapting models to analyze sensitive legal or healthcare documents.