Large language models (LLMs) demand substantial computational and memory resources, creating deployment challenges. Quantization-aware training (QAT) addresses these challenges by reducing model precision while maintaining performance. However, the scaling behavior of QAT, especially at 4-bit precision (W4A4), is not well understood. Existing QAT scaling laws often ignore key factors such as the number of training tokens and quantization granularity, which limits their applicability. This paper proposes a unified scaling law for QAT that models quantization error as a function of model size, training data volume, and quantization group size. Through 268 QAT experiments, we show that quantization error decreases as model size increases, but rises with more training tokens and coarser quantization granularity. To identify the sources of W4A4 quantization error, we decompose it into weight and activation components. Both components follow the overall trend of W4A4 quantization error, but with different sensitivities. Specifically, weight quantization error increases more rapidly with more training tokens. Further analysis shows that the activation quantization error in the FC2 layer, caused by outliers, is the primary bottleneck of W4A4 QAT quantization error. By applying mixed-precision quantization to address this bottleneck, we demonstrate that weight and activation quantization errors can converge to similar levels. Additionally, with more training data, weight quantization error eventually exceeds activation quantization error, suggesting that reducing weight quantization error is also important in such scenarios. These findings offer key insights for improving QAT research and development.
Deep neural networks have achieved state-of-the-art results in a wide range of applications, from natural language processing and computer vision to speech recognition. However, as tasks become increasingly complex, model sizes continue to grow, posing challenges in latency and memory efficiency. To meet these constraints, post-training quantization has emerged as a promising solution. In this paper, we propose a novel hardware-efficient quantization and inference scheme that exploits hardware advantages with minimal accuracy degradation. Specifically, we introduce a W4A8 scheme, where weights are quantized and stored using 4-bit integer precision, and inference computations are performed using 8-bit floating-point arithmetic, demonstrating significant speedups and improved memory utilization compared to 16-bit operations, applicable on various modern accelerators. To mitigate accuracy loss, we develop a novel quantization algorithm, dubbed Dual Precision Quantization (DPQ), that leverages the unique structure of our scheme without introducing additional inference overhead. Experimental results demonstrate improved performance (i.e., increased throughput) while maintaining tolerable accuracy degradation relative to the full-precision model.
Quantization methods are widely used to accelerate inference and streamline the deployment of large language models (LLMs). While prior research has extensively investigated the degradation of various LLM capabilities due to quantization, its effects on model explainability and interpretability, which are crucial for understanding decision-making processes, remain unexplored. To address this gap, we conduct comprehensive experiments using three common quantization techniques at distinct bit widths, in conjunction with two explainability methods, counterfactual examples and natural language explanations, as well as two interpretability approaches, knowledge memorization analysis and latent multi-hop reasoning analysis. We complement our analysis with a thorough user study, evaluating selected explainability methods. Our findings reveal that, depending on the configuration, quantization can significantly impact model explainability and interpretability. Notably, the direction of this effect is not consistent, as it strongly depends on (1) the quantization method, (2) the explainability or interpretability approach, and (3) the evaluation protocol. In some settings, human evaluation shows that quantization degrades explainability, while in others, it even leads to improvements. Our work serves as a cautionary tale, demonstrating that quantization can unpredictably affect model transparency. This insight has important implications for deploying LLMs in applications where transparency is a critical requirement.
Large Language Models (LLMs) have driven significant progress, yet their growing parameter counts and context windows incur prohibitive compute, energy, and monetary costs. We introduce EfficientLLM, a novel benchmark and the first comprehensive empirical study evaluating efficiency techniques for LLMs at scale. Conducted on a production-class cluster (48xGH200, 8xH200 GPUs), our study systematically explores three key axes: (1) architecture pretraining (efficient attention variants: MQA, GQA, MLA, NSA; sparse Mixture-of-Experts (MoE)), (2) fine-tuning (parameter-efficient methods: LoRA, RSLoRA, DoRA), and (3) inference (quantization methods: int4, float16). We define six fine-grained metrics (Memory Utilization, Compute Utilization, Latency, Throughput, Energy Consumption, Compression Rate) to capture hardware saturation, latency-throughput balance, and carbon cost. Evaluating over 100 model-technique pairs (0.5B-72B parameters), we derive three core insights: (i) Efficiency involves quantifiable trade-offs: no single method is universally optimal; e.g., MoE reduces FLOPs and improves accuracy but increases VRAM by 40%, while int4 quantization cuts memory/energy by up to 3.9x at a 3-5% accuracy drop. (ii) Optima are task- and scale-dependent: MQA offers optimal memory-latency trade-offs for constrained devices, MLA achieves lowest perplexity for quality-critical tasks, and RSLoRA surpasses LoRA efficiency only beyond 14B parameters. (iii) Techniques generalize across modalities: we extend evaluations to Large Vision Models (Stable Diffusion 3.5, Wan 2.1) and Vision-Language Models (Qwen2.5-VL), confirming effective transferability. By open-sourcing datasets, evaluation pipelines, and leaderboards, EfficientLLM provides essential guidance for researchers and engineers navigating the efficiency-performance landscape of next-generation foundation models.
Modern deep neural networks exhibit heterogeneity across numerous layers of various types such as residuals, multi-head attention, etc., due to varying structures (dimensions, activation functions, etc.), distinct representation characteristics, which impact predictions. We develop a general layer-wise quantization framework with tight variance and code-length bounds, adapting to the heterogeneities over the course of training. We then apply a new layer-wise quantization technique within distributed variational inequalities (VIs), proposing a novel Quantized Optimistic Dual Averaging (QODA) algorithm with adaptive learning rates, which achieves competitive convergence rates for monotone VIs. We empirically show that QODA achieves up to a $150\%$ speedup over the baselines in end-to-end training time for training Wasserstein GAN on $12+$ GPUs.
The exponential growth of visual data in digital communications has intensified the need for efficient compression techniques that balance rate-distortion performance with computational feasibility. While recent neural compression approaches have shown promise, they still struggle with fundamental challenges: preserving perceptual quality at high compression ratios, computational efficiency, and adaptability to diverse visual content. This paper introduces GANCompress, a novel neural compression framework that synergistically combines Binary Spherical Quantization (BSQ) with Generative Adversarial Networks (GANs) to address these challenges. Our approach employs a transformer-based autoencoder with an enhanced BSQ bottleneck that projects latent representations onto a hypersphere, enabling efficient discretization with bounded quantization error. This is followed by a specialized GAN architecture incorporating frequency-domain attention and color consistency optimization. Experimental results demonstrate that GANCompress achieves substantial improvement in compression efficiency -- reducing file sizes by up to 100x with minimal visual distortion. Our method outperforms traditional codecs like H.264 by 12-15% in perceptual metrics while maintaining comparable PSNR/SSIM values, with 2.4x faster encoding and decoding speeds. On standard benchmarks including ImageNet-1k and COCO2017, GANCompress sets a new state-of-the-art, reducing FID from 0.72 to 0.41 (43% improvement) compared to previous methods while maintaining higher throughput. This work presents a significant advancement in neural compression technology with promising applications for real-time visual communication systems.
We introduce Qronos -- a new state-of-the-art post-training quantization algorithm that sequentially rounds and updates neural network weights. Qronos not only explicitly corrects errors due to both weight and activation quantization, but also errors resulting from quantizing previous layers. Our iterative algorithm is based on an interpretable and disciplined optimization framework that subsumes and surpasses existing data-driven approaches. At each step, Qronos alternates between error correction and diffusion via optimal update rules. Importantly, we prove that Qronos admits an efficient implementation that uses the Cholesky decomposition for solving least-squares problems. We also demonstrate that Qronos is compatible with existing transformation techniques such as Hadamard-based incoherence processing and weight-activation scaling equalization, among others. We evaluate Qronos using recent autoregressive language generation models in the Llama3 family; Qronos consistently outperforms previous state-of-the-art adaptive rounding methods when quantizing the weights, activations, and/or KV caches.
Peak detection is a fundamental task in radar and has therefore been studied extensively in radar literature. However, Integrated Sensing and Communication (ISAC) systems for sixth generation (6G) cellular networks need to perform peak detection under hardware impairments and constraints imposed by the underlying system designed for communications. This paper presents a comparative study of classical Constant False Alarm Rate (CFAR)-based algorithms and a recently proposed Convolutional Neural Network (CNN)-based method for peak detection in ISAC radar images. To impose practical constraints of ISAC systems, we model the impact of hardware impairments, such as power amplifier nonlinearities and quantization noise. We perform extensive simulation campaigns focusing on multi-target detection under varying noise as well as on target separation in resolution-limited scenarios. The results show that CFAR detectors require approximate knowledge of the operating scenario and the use of window functions for reliable performance. The CNN, on the other hand, achieves high performance in all scenarios, but requires a preprocessing step for the input data.
In this paper, we consider discrete-time non-linear stochastic dynamical systems with additive process noise in which both the initial state and noise distributions are uncertain. Our goal is to quantify how the uncertainty in these distributions is propagated by the system dynamics for possibly infinite time steps. In particular, we model the uncertainty over input and noise as ambiguity sets of probability distributions close in the $\rho$-Wasserstein distance and aim to quantify how these sets evolve over time. Our approach relies on results from quantization theory, optimal transport, and stochastic optimization to construct ambiguity sets of distributions centered at mixture of Gaussian distributions that are guaranteed to contain the true sets for both finite and infinite prediction time horizons. We empirically evaluate the effectiveness of our framework in various benchmarks from the control and machine learning literature, showing how our approach can efficiently and formally quantify the uncertainty in linear and non-linear stochastic dynamical systems.
Using Resistive Random Access Memory (RRAM) crossbars in Computing-in-Memory (CIM) architectures offers a promising solution to overcome the von Neumann bottleneck. Due to non-idealities like cell variability, RRAM crossbars are often operated in binary mode, utilizing only two states: Low Resistive State (LRS) and High Resistive State (HRS). Binary Neural Networks (BNNs) and Ternary Neural Networks (TNNs) are well-suited for this hardware due to their efficient mapping. Existing software projects for RRAM-based CIM typically focus on only one aspect: compilation, simulation, or Design Space Exploration (DSE). Moreover, they often rely on classical 8 bit quantization. To address these limitations, we introduce CIM-Explorer, a modular toolkit for optimizing BNN and TNN inference on RRAM crossbars. CIM-Explorer includes an end-to-end compiler stack, multiple mapping options, and simulators, enabling a DSE flow for accuracy estimation across different crossbar parameters and mappings. CIM-Explorer can accompany the entire design process, from early accuracy estimation for specific crossbar parameters, to selecting an appropriate mapping, and compiling BNNs and TNNs for a finalized crossbar chip. In DSE case studies, we demonstrate the expected accuracy for various mappings and crossbar parameters. CIM-Explorer can be found on GitHub.
Existing works on Binary Neural Network (BNN) mainly focus on model's weights and activations while discarding considerations on the input raw data. This article introduces Generic Learned Thermometer (GLT), an encoding technique to improve input data representation for BNN, relying on learning non linear quantization thresholds. This technique consists in multiple data binarizations which can advantageously replace a conventional Analog to Digital Conversion (ADC) that uses natural binary coding. Additionally, we jointly propose a compact topology with light-weight grouped convolutions being trained thanks to block pruning and Knowledge Distillation (KD), aiming at reducing furthermore the model size so as its computational complexity. We show that GLT brings versatility to the BNN by intrinsically performing global tone mapping, enabling significant accuracy gains in practice (demonstrated by simulations on the STL-10 and VWW datasets). Moreover, when combining GLT with our proposed block-pruning technique, we successfully achieve lightweight (under 1Mb), fully-binarized models with limited accuracy degradation while being suitable for in-sensor always-on inference use cases.
Tackling complex optimization problems often relies on expert-designed heuristics, typically crafted through extensive trial and error. Recent advances demonstrate that large language models (LLMs), when integrated into well-designed evolutionary search frameworks, can autonomously discover high-performing heuristics at a fraction of the traditional cost. However, existing approaches predominantly rely on verbal guidance, i.e., manipulating the prompt generation process, to steer the evolution of heuristics, without adapting the underlying LLM. We propose a hybrid framework that combines verbal and numerical guidance, the latter achieved by fine-tuning the LLM via reinforcement learning based on the quality of generated heuristics. This joint optimization allows the LLM to co-evolve with the search process. Our method outperforms state-of-the-art (SOTA) baselines across various optimization tasks, running locally on a single 24GB GPU using a 7B model with INT4 quantization. It surpasses methods that rely solely on verbal guidance, even when those use significantly more powerful API-based models.
We introduce SLED, an alternative approach to speech language modeling by encoding speech waveforms into sequences of continuous latent representations and modeling them autoregressively using an energy distance objective. The energy distance offers an analytical measure of the distributional gap by contrasting simulated and target samples, enabling efficient training to capture the underlying continuous autoregressive distribution. By bypassing reliance on residual vector quantization, SLED avoids discretization errors and eliminates the need for the complicated hierarchical architectures common in existing speech language models. It simplifies the overall modeling pipeline while preserving the richness of speech information and maintaining inference efficiency. Empirical results demonstrate that SLED achieves strong performance in both zero-shot and streaming speech synthesis, showing its potential for broader applications in general-purpose speech language models.
The development of edge computing places critical demands on energy-efficient model deployment for multiple-input multiple-output (MIMO) detection tasks. Deploying deep unfolding models such as PGD-Nets and ADMM-Nets into resource-constrained edge devices using quantization methods is challenging. Existing quantization methods based on quantization aware training (QAT) suffer from performance degradation due to their reliance on parametric distribution assumption of activations and static quantization step sizes. To address these challenges, this paper proposes a novel kernel-based adaptive quantization (KAQ) framework for deep unfolding networks. By utilizing a joint kernel density estimation (KDE) and maximum mean discrepancy (MMD) approach to align activation distributions between full-precision and quantized models, the need for prior distribution assumptions is eliminated. Additionally, a dynamic step size updating method is introduced to adjust the quantization step size based on the channel conditions of wireless networks. Extensive simulations demonstrate that the accuracy of proposed KAQ framework outperforms traditional methods and successfully reduces the model's inference latency.
Quantization is essential for Neural Network (NN) compression, reducing model size and computational demands by using lower bit-width data types, though aggressive reduction often hampers accuracy. Mixed Precision (MP) mitigates this tradeoff by varying the numerical precision across network layers. This study focuses on automatically selecting an optimal MP configuration within Post-Training Quantization (PTQ) for inference. The first key contribution is a novel sensitivity metric derived from a first-order Taylor series expansion of the loss function as a function of quantization errors in weights and activations. This metric, based on the Mean Square Error (MSE) of the loss, is efficiently calculated per layer using high-precision forward and backward passes over a small calibration dataset. The metric is additive across layers, with low calibration memory overhead as weight optimization is unnecessary. The second contribution is an accurate hardware-aware method for predicting MP time gain by modeling it as additive for sequential sub-graphs. An algorithm partitions the model graph into sequential subgraphs, measuring time gain for each configuration using a few samples. After calibrating per-layer sensitivity and time gain, an Integer Programming (IP) problem is formulated to maximize time gain while keeping loss MSE below a set threshold. Memory gain and theoretical time gain based on Multiply and Accumulate (MAC) operations are also considered. Rigorous experiments on the Intel Gaudi 2 accelerator validate the approach on several Large Language Models (LLMs).
Multi-frame video enhancement tasks aim to improve the spatial and temporal resolution and quality of video sequences by leveraging temporal information from multiple frames, which are widely used in streaming video processing, surveillance, and generation. Although numerous Transformer-based enhancement methods have achieved impressive performance, their computational and memory demands hinder deployment on edge devices. Quantization offers a practical solution by reducing the bit-width of weights and activations to improve efficiency. However, directly applying existing quantization methods to video enhancement tasks often leads to significant performance degradation and loss of fine details. This stems from two limitations: (a) inability to allocate varying representational capacity across frames, which results in suboptimal dynamic range adaptation; (b) over-reliance on full-precision teachers, which limits the learning of low-bit student models. To tackle these challenges, we propose a novel quantization method for video enhancement: Progressive Multi-Frame Quantization for Video Enhancement (PMQ-VE). This framework features a coarse-to-fine two-stage process: Backtracking-based Multi-Frame Quantization (BMFQ) and Progressive Multi-Teacher Distillation (PMTD). BMFQ utilizes a percentile-based initialization and iterative search with pruning and backtracking for robust clipping bounds. PMTD employs a progressive distillation strategy with both full-precision and multiple high-bit (INT) teachers to enhance low-bit models' capacity and quality. Extensive experiments demonstrate that our method outperforms existing approaches, achieving state-of-the-art performance across multiple tasks and benchmarks.The code will be made publicly available at: https://github.com/xiaoBIGfeng/PMQ-VE.
As the size of large language models grows exponentially, GPU memory has become a bottleneck for adapting these models to downstream tasks. In this paper, we aim to push the limits of memory-efficient training by minimizing memory usage on model weights, gradients, and optimizer states, within a unified framework. Our idea is to eliminate both gradients and optimizer states using zeroth-order optimization, which approximates gradients by perturbing weights during forward passes to identify gradient directions. To minimize memory usage on weights, we employ model quantization, e.g., converting from bfloat16 to int4. However, directly applying zeroth-order optimization to quantized weights is infeasible due to the precision gap between discrete weights and continuous gradients, which would otherwise require de-quantization and re-quantization. To overcome this challenge, we propose Quantized Zeroth-order Optimization (QZO), a novel approach that perturbs the continuous quantization scale for gradient estimation and uses a directional derivative clipping method to stabilize training. QZO is orthogonal to both scalar-based and codebook-based post-training quantization methods. Compared to full-parameter fine-tuning in bfloat16, QZO can reduce the total memory cost by more than 18$\times$ for 4-bit LLMs, and enables fine-tuning Llama-2-13B and Stable Diffusion 3.5 Large within a single 24GB GPU.
Hierarchical data arise in countless domains, from biological taxonomies and organizational charts to legal codes and knowledge graphs. Residual Quantization (RQ) is widely used to generate discrete, multitoken representations for such data by iteratively quantizing residuals in a multilevel codebook. However, its reliance on Euclidean geometry can introduce fundamental mismatches that hinder modeling of hierarchical branching, necessary for faithful representation of hierarchical data. In this work, we propose Hyperbolic Residual Quantization (HRQ), which embeds data natively in a hyperbolic manifold and performs residual quantization using hyperbolic operations and distance metrics. By adapting the embedding network, residual computation, and distance metric to hyperbolic geometry, HRQ imparts an inductive bias that aligns naturally with hierarchical branching. We claim that HRQ in comparison to RQ can generate more useful for downstream tasks discrete hierarchical representations for data with latent hierarchies. We evaluate HRQ on two tasks: supervised hierarchy modeling using WordNet hypernym trees, where the model is supervised to learn the latent hierarchy - and hierarchy discovery, where, while latent hierarchy exists in the data, the model is not directly trained or evaluated on a task related to the hierarchy. Across both scenarios, HRQ hierarchical tokens yield better performance on downstream tasks compared to Euclidean RQ with gains of up to $20\%$ for the hierarchy modeling task. Our results demonstrate that integrating hyperbolic geometry into discrete representation learning substantially enhances the ability to capture latent hierarchies.
Human motion synthesis in complex scenes presents a fundamental challenge, extending beyond conventional Text-to-Motion tasks by requiring the integration of diverse modalities such as static environments, movable objects, natural language prompts, and spatial waypoints. Existing language-conditioned motion models often struggle with scene-aware motion generation due to limitations in motion tokenization, which leads to information loss and fails to capture the continuous, context-dependent nature of 3D human movement. To address these issues, we propose UniHM, a unified motion language model that leverages diffusion-based generation for synthesizing scene-aware human motion. UniHM is the first framework to support both Text-to-Motion and Text-to-Human-Object Interaction (HOI) in complex 3D scenes. Our approach introduces three key contributions: (1) a mixed-motion representation that fuses continuous 6DoF motion with discrete local motion tokens to improve motion realism; (2) a novel Look-Up-Free Quantization VAE (LFQ-VAE) that surpasses traditional VQ-VAEs in both reconstruction accuracy and generative performance; and (3) an enriched version of the Lingo dataset augmented with HumanML3D annotations, providing stronger supervision for scene-specific motion learning. Experimental results demonstrate that UniHM achieves comparative performance on the OMOMO benchmark for text-to-HOI synthesis and yields competitive results on HumanML3D for general text-conditioned motion generation.
Large language models have demonstrated remarkable performance; however, their massive parameter counts make deployment highly expensive. Low-rank approximation offers a promising compression solution, yet existing approaches have two main limitations: (1) They focus on minimizing the output error of individual linear layers, without considering the architectural characteristics of Transformers, and (2) they decompose a large weight matrix into two small low-rank matrices. Consequently, these methods often fall short compared to other compression techniques like pruning and quantization, and introduce runtime overhead such as the extra GEMM kernel launches for decomposed small matrices. To address these limitations, we propose $\tt A^\tt 3$, a post-training low-rank approximation framework. $\tt A^\tt 3$ splits a Transformer layer into three functional components, namely $\tt QK$, $\tt OV$, and $\tt MLP$. For each component, $\tt A^\tt 3$ provides an analytical solution that reduces the hidden dimension size inside each component while minimizing the component's functional loss ($\it i.e.$, error in attention scores, attention outputs, and MLP outputs). This approach directly reduces model sizes, KV cache sizes, and FLOPs without introducing any runtime overheads. In addition, it provides a new narrative in advancing the optimization problem from singular linear layer loss optimization toward improved end-to-end performance. Through extensive experiments, we show that $\tt A^\tt 3$ maintains superior performance compared to SoTAs. For example, under the same reduction budget in computation and memory, our low-rank approximated LLaMA 3.1-70B achieves a perplexity of 4.69 on WikiText-2, outperforming the previous SoTA's 7.87 by 3.18. We also demonstrate the versatility of $\tt A^\tt 3$, including KV cache compression, quantization, and mixed-rank assignments for enhanced performance.