Federated learning (FL) enables collaborative training across distributed clients without sharing raw data, often at the cost of substantial communication overhead induced by transmitting high-dimensional model updates. This overhead can be alleviated by having the clients quantize their model updates, with dithered lattice quantizers identified as an attractive scheme due to its structural simplicity and convergence-preserving properties. However, existing lattice-based FL schemes typically rely on a fixed quantization rule, which is suboptimal in heterogeneous and dynamic environments where the model updates distribution varies across users and training rounds. In this work, we propose Online Learned Adaptive Lattices (OLALa), a heterogeneous FL framework where each client can adjust its quantizer online using lightweight local computations. We first derive convergence guarantees for FL with non-fixed lattice quantizers and show that proper lattice adaptation can tighten the convergence bound. Then, we design an online learning algorithm that enables clients to tune their quantizers throughout the FL process while exchanging only a compact set of quantization parameters. Numerical experiments demonstrate that OLALa consistently improves learning performance under various quantization rates, outperforming conventional fixed-codebook and non-adaptive schemes.
This paper presents a comprehensive evaluation of skin color measurement methods from dermatoscopic images using a synthetic dataset (S-SYNTH) with controlled ground-truth melanin content, lesion shapes, hair models, and 18 distinct lighting conditions. This allows for rigorous assessment of the robustness and invariance to lighting conditions. We assess four classes of image colorimetry approaches: segmentation-based, patch-based, color quantization, and neural networks. We use these methods to estimate the Individual Typology Angle (ITA) and Fitzpatrick types from dermatoscopic images. Our results show that segmentation-based and color quantization methods yield robust, lighting-invariant estimates, whereas patch-based approaches exhibit significant lighting-dependent biases that require calibration. Furthermore, neural network models, particularly when combined with heavy blurring to reduce overfitting, can provide light-invariant Fitzpatrick predictions, although their generalization to real-world images remains unverified. We conclude with practical recommendations for designing fair and reliable skin color estimation methods.
The ever-increasing computational demands and deployment costs of large language models (LLMs) have spurred numerous compressing methods. Compared to quantization and unstructured pruning, SVD compression offers superior hardware compatibility and theoretical guarantees. However, existing SVD-based methods focus on the overall discrepancy between the original and compressed matrices while overlooking the protection of critical components within the matrix, which leads to inferior performance in the compressed models. This paper proposes a dual-level importance protection mechanism to enhance SVD-based compression methods: (1) local importance protection: preserving the most critical singular vectors within each weight matrix through channel-weighted data whitening; and (2) global importance protection: enabling less important layers to bear a greater portion of the compression burden through either a heuristic or optimization-based approach, thereby minimizing the impact of compression on critical layers. Extensive experiments demonstrate that DipSVD outperforms existing SVD-based compression approaches across multiple benchmarks, achieving superior model performance especially at high model compression ratios.
Quantized large language models (LLMs) have gained increasing attention and significance for enabling deployment in resource-constrained environments. However, emerging studies on a few calibration dataset-free quantization methods suggest that quantization may compromise the safety capabilities of LLMs, underscoring the urgent need for systematic safety evaluations and effective mitigation strategies. In this paper, we present comprehensive safety evaluations across various mainstream quantization techniques and diverse calibration datasets, utilizing widely accepted safety benchmarks. To address the identified safety vulnerabilities, we propose a quantization-aware safety patching framework, Q-resafe, to efficiently restore the safety capabilities of quantized LLMs while minimizing any adverse impact on utility. Extensive experimental results demonstrate that Q-resafe successfully re-aligns the safety of quantized LLMs with their pre-quantization counterparts, even under challenging evaluation scenarios. Project page is available at: https://github.com/Thecommonirin/Qresafe.
Uncovering emergent concepts across transformer layers remains a significant challenge because the residual stream linearly mixes and duplicates information, obscuring how features evolve within large language models. Current research efforts primarily inspect neural representations at single layers, thereby overlooking this cross-layer superposition and the redundancy it introduces. These representations are typically either analyzed directly for activation patterns or passed to probing classifiers that map them to a limited set of predefined concepts. To address these limitations, we propose \gls{clvqvae}, a framework that uses vector quantization to map representations across layers and in the process collapse duplicated residual-stream features into compact, interpretable concept vectors. Our approach uniquely combines top-$k$ temperature-based sampling during quantization with EMA codebook updates, providing controlled exploration of the discrete latent space while maintaining code-book diversity. We further enhance the framework with scaled-spherical k-means++ for codebook initialization, which clusters by directional similarity rather than magnitude, better aligning with semantic structure in word embedding space.
Analog In-memory Computing (AIMC) has emerged as a highly efficient paradigm for accelerating Deep Neural Networks (DNNs), offering significant energy and latency benefits over conventional digital hardware. However, state-of-the-art neural networks are not inherently designed for AIMC, as they fail to account for its unique non-idealities. Neural Architecture Search (NAS) is thus needed to systematically discover neural architectures optimized explicitly for AIMC constraints. However, comparing NAS methodologies and extracting insights about robust architectures for AIMC requires a dedicated NAS benchmark that explicitly accounts for AIMC-specific hardware non-idealities. To address this, we introduce AnalogNAS-Bench, the first NAS benchmark tailored specifically for AIMC. Our study reveals three key insights: (1) standard quantization techniques fail to capture AIMC-specific noises, (2) robust architectures tend to feature wider and branched blocks, (3) skip connections improve resilience to temporal drift noise. These insights highlight the limitations of current NAS benchmarks for AIMC and pave the way for future analog-aware NAS. All the implementations used in this paper can be found at https://github.com/IBM/analog-nas/tree/main/analognasbench.
This paper presents J3DAI, a tiny deep neural network-based hardware accelerator for a 3-layer 3D-stacked CMOS image sensor featuring an artificial intelligence (AI) chip integrating a Deep Neural Network (DNN)-based accelerator. The DNN accelerator is designed to efficiently perform neural network tasks such as image classification and segmentation. This paper focuses on the digital system of J3DAI, highlighting its Performance-Power-Area (PPA) characteristics and showcasing advanced edge AI capabilities on a CMOS image sensor. To support hardware, we utilized the Aidge comprehensive software framework, which enables the programming of both the host processor and the DNN accelerator. Aidge supports post-training quantization, significantly reducing memory footprint and computational complexity, making it crucial for deploying models on resource-constrained hardware like J3DAI. Our experimental results demonstrate the versatility and efficiency of this innovative design in the field of edge AI, showcasing its potential to handle both simple and computationally intensive tasks. Future work will focus on further optimizing the architecture and exploring new applications to fully leverage the capabilities of J3DAI. As edge AI continues to grow in importance, innovations like J3DAI will play a crucial role in enabling real-time, low-latency, and energy-efficient AI processing at the edge.
Quantization has been widely studied as an effective technique for reducing the memory requirement of large language models (LLMs), potentially improving the latency time as well. Utilizing the characteristic of rotational invariance of transformer, we propose the rotation-based saliency-aware weight quantization (ROSAQ), which identifies salient channels in the projection feature space, not in the original feature space, where the projected "principal" dimensions are naturally considered as "salient" features. The proposed ROSAQ consists of 1) PCA-based projection, which first performs principal component analysis (PCA) on a calibration set and transforms via the PCA projection, 2) Salient channel dentification, which selects dimensions corresponding to the K-largest eigenvalues as salient channels, and 3) Saliency-aware quantization with mixed-precision, which uses FP16 for salient dimensions and INT3/4 for other dimensions. Experiment results show that ROSAQ shows improvements over the baseline saliency-aware quantization on the original feature space and other existing quantization methods. With kernel fusion, ROSAQ presents about 2.3x speed up over FP16 implementation in generating 256 tokens with a batch size of 64.
Diffusion models are well known for their ability to generate a high-fidelity image for an input prompt through an iterative denoising process. Unfortunately, the high fidelity also comes at a high computational cost due the inherently sequential generative process. In this work, we seek to optimally balance quality and computational cost, and propose a framework to allow the amount of computation to vary for each prompt, depending on its complexity. Each prompt is automatically routed to the most appropriate text-to-image generation function, which may correspond to a distinct number of denoising steps of a diffusion model, or a disparate, independent text-to-image model. Unlike uniform cost reduction techniques (e.g., distillation, model quantization), our approach achieves the optimal trade-off by learning to reserve expensive choices (e.g., 100+ denoising steps) only for a few complex prompts, and employ more economical choices (e.g., small distilled model) for less sophisticated prompts. We empirically demonstrate on COCO and DiffusionDB that by learning to route to nine already-trained text-to-image models, our approach is able to deliver an average quality that is higher than that achievable by any of these models alone.
This paper presents a general-purpose video super-resolution (VSR) method, dubbed VSR-HE, specifically designed to enhance the perceptual quality of compressed content. Targeting scenarios characterized by heavy compression, the method upscales low-resolution videos by a ratio of four, from 180p to 720p or from 270p to 1080p. VSR-HE adopts hierarchical encoding transformer blocks and has been sophisticatedly optimized to eliminate a wide range of compression artifacts commonly introduced by H.265/HEVC encoding across various quantization parameter (QP) levels. To ensure robustness and generalization, the model is trained and evaluated under diverse compression settings, allowing it to effectively restore fine-grained details and preserve visual fidelity. The proposed VSR-HE has been officially submitted to the ICME 2025 Grand Challenge on VSR for Video Conferencing (Team BVI-VSR), under both the Track 1 (General-Purpose Real-World Video Content) and Track 2 (Talking Head Videos).
Deploying large language models (LLMs) often faces challenges from substantial memory and computational costs. Quantization offers a solution, yet performance degradation in the sub-1-bit regime remains particularly difficult. This paper introduces LittleBit, a novel method for extreme LLM compression. It targets levels like 0.1 bits per weight (BPW), achieving nearly 31$\times$ memory reduction, e.g., Llama2-13B to under 0.9 GB. LittleBit represents weights in a low-rank form using latent matrix factorization, subsequently binarizing these factors. To counteract information loss from this extreme precision, it integrates a multi-scale compensation mechanism. This includes row, column, and an additional latent dimension that learns per-rank importance. Two key contributions enable effective training: Dual Sign-Value-Independent Decomposition (Dual-SVID) for stable quantization-aware training (QAT) initialization, and integrated Residual Compensation to mitigate errors. Extensive experiments confirm LittleBit's superiority in sub-1-bit quantization: e.g., its 0.1 BPW performance on Llama2-7B surpasses the leading method's 0.7 BPW. This establishes a superior size-performance trade-off, with kernel-level benchmarks indicating potential for a 5$\times$ speedup compared to FP16. LittleBit paves the way for deploying powerful LLMs in resource-constrained environments.
Large multimodal Mixture-of-Experts (MoEs) effectively scale the model size to boost performance while maintaining fixed active parameters. However, previous works primarily utilized full-precision experts during sparse up-cycling. Despite they show superior performance on end tasks, the large amount of experts introduces higher memory footprint, which poses significant challenges for the deployment on edge devices. In this work, we propose MoTE, a scalable and memory-efficient approach to train Mixture-of-Ternary-Experts models from dense checkpoint. Instead of training fewer high-precision experts, we propose to train more low-precision experts during up-cycling. Specifically, we use the pre-trained FFN as a shared expert and train ternary routed experts with parameters in {-1, 0, 1}. Extensive experiments show that our approach has promising scaling trend along model size. MoTE achieves comparable performance to full-precision baseline MoE-LLaVA while offering lower memory footprint. Furthermore, our approach is compatible with post-training quantization methods and the advantage further amplifies when memory-constraint goes lower. Given the same amount of expert memory footprint of 3.4GB and combined with post-training quantization, MoTE outperforms MoE-LLaVA by a gain of 4.3% average accuracy on end tasks, demonstrating its effectiveness and potential for memory-constrained devices.
With the rapid growth of Internet services, recommendation systems play a central role in delivering personalized content. Faced with massive user requests and complex model architectures, the key challenge for real-time recommendation systems is how to reduce inference latency and increase system throughput without sacrificing recommendation quality. This paper addresses the high computational cost and resource bottlenecks of deep learning models in real-time settings by proposing a combined set of modeling- and system-level acceleration and optimization strategies. At the model level, we dramatically reduce parameter counts and compute requirements through lightweight network design, structured pruning, and weight quantization. At the system level, we integrate multiple heterogeneous compute platforms and high-performance inference libraries, and we design elastic inference scheduling and load-balancing mechanisms based on real-time load characteristics. Experiments show that, while maintaining the original recommendation accuracy, our methods cut latency to less than 30% of the baseline and more than double system throughput, offering a practical solution for deploying large-scale online recommendation services.
The rapid evolution of large language models (LLMs), driven by growing parameter scales, adoption of mixture-of-experts (MoE) architectures, and expanding context lengths, imposes unprecedented demands on AI infrastructure. Traditional AI clusters face limitations in compute intensity, memory bandwidth, inter-chip communication, and latency, compounded by variable workloads and strict service-level objectives. Addressing these issues requires fundamentally redesigned hardware-software integration. This paper introduces Huawei CloudMatrix, a next-generation AI datacenter architecture, realized in the production-grade CloudMatrix384 supernode. It integrates 384 Ascend 910 NPUs and 192 Kunpeng CPUs interconnected via an ultra-high-bandwidth Unified Bus (UB) network, enabling direct all-to-all communication and dynamic pooling of resources. These features optimize performance for communication-intensive operations, such as large-scale MoE expert parallelism and distributed key-value cache access. To fully leverage CloudMatrix384, we propose CloudMatrix-Infer, an advanced LLM serving solution incorporating three core innovations: a peer-to-peer serving architecture that independently scales prefill, decode, and caching; a large-scale expert parallelism strategy supporting EP320 via efficient UB-based token dispatch; and hardware-aware optimizations including specialized operators, microbatch-based pipelining, and INT8 quantization. Evaluation with the DeepSeek-R1 model shows CloudMatrix-Infer achieves state-of-the-art efficiency: prefill throughput of 6,688 tokens/s per NPU and decode throughput of 1,943 tokens/s per NPU (<50 ms TPOT). It effectively balances throughput and latency, sustaining 538 tokens/s per NPU even under stringent 15 ms latency constraints, while INT8 quantization maintains model accuracy across benchmarks.
Mixture-of-Experts (MoE) models have emerged as a cornerstone of large-scale deep learning by efficiently distributing computation and enhancing performance. However, their unique architecture-characterized by sparse expert activation and dynamic routing mechanisms-introduces inherent complexities that challenge conventional quantization techniques. Existing post-training quantization (PTQ) methods struggle to address activation outliers, router consistency and sparse expert calibration, leading to significant performance degradation. To bridge this gap, we propose EAQuant, a novel PTQ framework tailored for MoE architectures. Our method systematically tackles these challenges through three key innovations: (1) expert-aware smoothing aggregation to suppress activation outliers and stabilize quantization, (2) router logits distribution alignment to preserve expert selection consistency post-quantization, and (3) expert-level calibration data balance to optimize sparsely activated experts. Extensive experiments across W4A4 and extreme W3A4 quantization configurations demonstrate that EAQuant significantly outperforms existing methods, achieving average score improvements of 1.15 - 2.28% across three diverse MoE architectures, with particularly pronounced gains in reasoning tasks and robust performance retention under aggressive quantization. By integrating these innovations, EAQuant establishes a new state-of-the-art for high-precision, efficient MoE model compression. Our code is available at https://github.com/darren-fzq/EAQuant.
Multi-person pose estimation (MPPE) estimates keypoints for all individuals present in an image. MPPE is a fundamental task for several applications in computer vision and virtual reality. Unfortunately, there are currently no transformer-based models that can perform MPPE in real time. The paper presents a family of transformer-based models capable of performing multi-person 2D pose estimation in real-time. Our approach utilizes a modified decoder architecture and keypoint similarity metrics to generate both positive and negative queries, thereby enhancing the quality of the selected queries within the architecture. Compared to state-of-the-art models, our proposed models train much faster, using 5 to 10 times fewer epochs, with competitive inference times without requiring quantization libraries to speed up the model. Furthermore, our proposed models provide competitive results or outperform alternative models, often using significantly fewer parameters.
Post-training quantization (PTQ) has stood out as a cost-effective and promising model compression paradigm in recent years, as it avoids computationally intensive model retraining. Nevertheless, current PTQ methods for Vision Transformers (ViTs) still suffer from significant accuracy degradation, especially under low-bit quantization. To address these shortcomings, we analyze the prevailing Hessian-guided quantization loss, and uncover certain limitations of conventional Hessian approximations. By following the block-wise reconstruction framework, we propose a novel PTQ method for ViTs, dubbed FIMA-Q. Specifically, we firstly establish the connection between KL divergence and FIM, which enables fast computation of the quantization loss during reconstruction. We further propose an efficient FIM approximation method, namely DPLR-FIM, by employing the diagonal plus low-rank principle, and formulate the ultimate quantization loss. Our extensive experiments, conducted across various vision tasks with representative ViT-based architectures on public datasets, demonstrate that our method substantially promotes the accuracy compared to the state-of-the-art approaches, especially in the case of low-bit quantization. The source code is available at https://github.com/ShiheWang/FIMA-Q.
Vision Transformers (ViTs) are essential in computer vision but are computationally intensive, too. Model quantization, particularly to low bit-widths like 4-bit, aims to alleviate this difficulty, yet existing Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT) methods exhibit significant limitations. PTQ often incurs substantial accuracy drop, while QAT achieves high accuracy but suffers from prohibitive computational costs, limited generalization to downstream tasks, training instability, and lacking of open-source codebase. To address these challenges, this paper introduces General, Practical, and Lightning Quantization (GPLQ), a novel framework designed for efficient and effective ViT quantization. GPLQ is founded on two key empirical insights: the paramount importance of activation quantization and the necessity of preserving the model's original optimization ``basin'' to maintain generalization. Consequently, GPLQ employs a sequential ``activation-first, weights-later'' strategy. Stage 1 keeps weights in FP32 while quantizing activations with a feature mimicking loss in only 1 epoch to keep it stay in the same ``basin'', thereby preserving generalization. Stage 2 quantizes weights using a PTQ method. As a result, GPLQ is 100x faster than existing QAT methods, lowers memory footprint to levels even below FP32 training, and achieves 4-bit model performance that is highly competitive with FP32 models in terms of both accuracy on ImageNet and generalization to diverse downstream tasks, including fine-grained visual classification and object detection. We will release an easy-to-use open-source toolkit supporting multiple vision tasks.
Large Language Models (LLMs) have significant impact on the healthcare scenarios but remain prohibitively large for deployment in real-time, resource-constrained environments such as edge devices. In this work, we introduce a novel medical assistant system, optimized through our general-purpose compression framework, which tailors Large Language Models (LLMs) for deployment in specialized domains. By measuring neuron saliency on domain-specific data, our method can aggressively prune irrelevant neurons, reducing model size while preserving performance. Following pruning, we apply post-training quantization to further reduce the memory footprint, and evaluate the compressed model across medical benchmarks including MedMCQA, MedQA, and PubMedQA. We also deploy the 50\% compressed Gemma and the 67\% compressed LLaMA3 models on Jetson Orin Nano (18.7W peak) and Raspberry Pi 5 (6.3W peak), achieving real-time, energy-efficient inference under hardware constraints.
Complexity of Neural Networks is increasing rapidly due to the massive increase in model parameters. Specifically, in Large Language Models (LLMs), the number of model parameters has grown exponentially in the past few years, for example, from 1.5 billion parameters in GPT2 to 175 billion in GPT3. This raises a significant challenge for implementation, especially for Edge devices where memory and processing power are very limited. In this work, we investigate reducing LLM complexity with special type of quantization, power of two (PoT), for linear layers weights and transformer tables. PoT not only provides memory reduction but more importantly provides significant computational reduction through converting multiplication to bit shifting. We obtained preliminary results of PoT quantization on Nano-GPT implementation using Shakespeare dataset. We then extended results to 124-M GPT-2 model. The PoT quantization results are shown to be very promising with cross entropy loss degradation $\approx$[1.3-0.88] with number of bits range [4-6] to represent power levels.