Recent speaker verification (SV) systems have shown a trend toward adopting deeper speaker embedding extractors. Although deeper and larger neural networks can significantly improve performance, their substantial memory requirements hinder training on consumer GPUs. In this paper, we explore a memory-efficient training strategy for deep speaker embedding learning in resource-constrained scenarios. Firstly, we conduct a systematic analysis of GPU memory allocation during SV system training. Empirical observations show that activations and optimizer states are the main sources of memory consumption. For activations, we design two types of reversible neural networks which eliminate the need to store intermediate activations during back-propagation, thereby significantly reducing memory usage without performance loss. For optimizer states, we introduce a dynamic quantization approach that replaces the original 32-bit floating-point values with a dynamic tree-based 8-bit data type. Experimental results on VoxCeleb demonstrate that the reversible variants of ResNets and DF-ResNets can perform training without the need to cache activations in GPU memory. In addition, the 8-bit versions of SGD and Adam save 75% of memory costs while maintaining performance compared to their 32-bit counterparts. Finally, a detailed comparison of memory usage and performance indicates that our proposed models achieve up to 16.2x memory savings, with nearly identical parameters and performance compared to the vanilla systems. In contrast to the previous need for multiple high-end GPUs such as the A100, we can effectively train deep speaker embedding extractors with just one or two consumer-level 2080Ti GPUs.
Low-rank adaptation (LoRA) has become the dominant method for parameter-efficient LLM fine-tuning, with LoRA-based quantization error compensation (LQEC) emerging as a powerful tool for recovering accuracy in compressed LLMs. However, LQEC has underperformed in sub-4-bit scenarios, with no prior investigation into understanding this limitation. We propose RILQ (Rank-Insensitive LoRA-based Quantization Error Compensation) to understand fundamental limitation and boost 2-bit LLM accuracy. Based on rank analysis revealing model-wise activation discrepancy loss's rank-insensitive nature, RILQ employs this loss to adjust adapters cooperatively across layers, enabling robust error compensation with low-rank adapters. Evaluations on LLaMA-2 and LLaMA-3 demonstrate RILQ's consistent improvements in 2-bit quantized inference across various state-of-the-art quantizers and enhanced accuracy in task-specific fine-tuning. RILQ maintains computational efficiency comparable to existing LoRA methods, enabling adapter-merged weight-quantized LLM inference with significantly enhanced accuracy, making it a promising approach for boosting 2-bit LLM performance. Our code is available at https://github.com/aiha-lab/RILQ.
Energy efficiency of Convolutional Neural Networks (CNNs) has become an important area of research, with various strategies being developed to minimize the power consumption of these models. Previous efforts, including techniques like model pruning, quantization, and hardware optimization, have made significant strides in this direction. However, there remains a need for more effective on device AI solutions that balance energy efficiency with model performance. In this paper, we propose a novel approach to reduce the energy requirements of inference of CNNs. Our methodology employs two small Complementary CNNs that collaborate with each other by covering each other's "weaknesses" in predictions. If the confidence for a prediction of the first CNN is considered low, the second CNN is invoked with the aim of producing a higher confidence prediction. This dual-CNN setup significantly reduces energy consumption compared to using a single large deep CNN. Additionally, we propose a memory component that retains previous classifications for identical inputs, bypassing the need to re-invoke the CNNs for the same input, further saving energy. Our experiments on a Jetson Nano computer demonstrate an energy reduction of up to 85.8% achieved on modified datasets where each sample was duplicated once. These findings indicate that leveraging a complementary CNN pair along with a memory component effectively reduces inference energy while maintaining high accuracy.
Time series analysis is a fundamental data mining task that supervised training methods based on empirical risk minimization have proven their effectiveness on specific tasks and datasets. However, the acquisition of well-annotated data is costly and a large amount of unlabeled series data is under-utilized. Due to distributional shifts across various domains and different patterns of interest across multiple tasks. The problem of cross-domain multi-task migration of time series remains a significant challenge. To address these problems, this paper proposes a novel cross-domain pretraining method based on Wave Quantization (termed as WQ4TS), which can be combined with any advanced time series model and applied to multiple downstream tasks. Specifically, we transfer the time series data from different domains into a common spectral latent space, and enable the model to learn the temporal pattern knowledge of different domains directly from the common space and utilize it for the inference of downstream tasks, thereby mitigating the challenge of heterogeneous cross-domains migration. The establishment of spectral latent space brings at least three benefits, cross-domain migration capability thus adapting to zero- and few-shot scenarios without relying on priori knowledge of the dataset, general compatible cross-domain migration framework without changing the existing model structure, and robust modeling capability thus achieving SOTA results in multiple downstream tasks. To demonstrate the effectiveness of the proposed approach, we conduct extensive experiments including three important tasks: forecasting, imputation, and classification. And three common real-world data scenarios are simulated: full-data, few-shot, and zero-shot. The proposed WQ4TS achieves the best performance on 87.5% of all tasks, and the average improvement of the metrics on all the tasks is up to 34.7%.
Vision tokenizers have gained a lot of attraction due to their scalability and compactness; previous works depend on old-school GAN-based hyperparameters, biased comparisons, and a lack of comprehensive analysis of the scaling behaviours. To tackle those issues, we introduce Grouped Spherical Quantization (GSQ), featuring spherical codebook initialization and lookup regularization to constrain codebook latent to a spherical surface. Our empirical analysis of image tokenizer training strategies demonstrates that GSQ-GAN achieves superior reconstruction quality over state-of-the-art methods with fewer training iterations, providing a solid foundation for scaling studies. Building on this, we systematically examine the scaling behaviours of GSQ, specifically in latent dimensionality, codebook size, and compression ratios, and their impact on model performance. Our findings reveal distinct behaviours at high and low spatial compression levels, underscoring challenges in representing high-dimensional latent spaces. We show that GSQ can restructure high-dimensional latent into compact, low-dimensional spaces, thus enabling efficient scaling with improved quality. As a result, GSQ-GAN achieves a 16x down-sampling with a reconstruction FID (rFID) of 0.50.
This paper analyzes the performance of Small Language Models (SLMs) and Vision Language Models (VLMs) and evaluates the trade-off between model performance and carbon emissions across 4 essential tasks: Image Captioning, Visual Question Answering (VQA), Dialogue Summarization and Text-to-SQL conversion. Various SLMs and VLMs belonging to the Qwen and LLaMA architecture family are chosen and variants based on model size in terms of the number of parameters, quantization level and fine-tuning parameters are evaluated. The model variant's performance and carbon emissions are calculated. To quantify the trade-off between model performance and carbon emissions, we introduce a novel metric called CEGI (Carbon Efficient Gain Index). This metric represents the carbon emission per unit percentage gain per million trainable parameters . This metric provides a normalized measure to compare model's efficiency in terms of performance improvement relative to their environmental cost. The experiment's outcome demonstrates that fine-tuning SLMs and VLMs can achieve performance levels comparable to Large Language Models (LLMs) while producing significantly less carbon emissions. Our findings suggest that the marginal gains in accuracy from larger models do not justify the substantial increase in carbon emissions. Leveraging lower-bit quantization levels, the proposed metric further enhances energy efficiency without compromising performance. This study highlights balancing high performance and environmental sustainability. It offers a valuable metric for selecting models suitable for environmentally-friendly AI development.
Existing vector quantization (VQ) methods struggle with scalability, largely attributed to the instability of the codebook that undergoes partial updates during training. The codebook is prone to collapse as utilization decreases, due to the progressively widening distribution gap between non-activated codes and visual features. To solve the problem, we propose Index Backpropagation Quantization (IBQ), a new VQ method for the joint optimization of all codebook embeddings and the visual encoder. Applying a straight-through estimator on the one-hot categorical distribution between the encoded feature and codebook, all codes are differentiable and maintain a consistent latent space with the visual encoder. IBQ enables scalable training of visual tokenizers and, for the first time, achieves a large-scale codebook ($2^{18}$) with high dimension ($256$) and high utilization. Experiments on the standard ImageNet benchmark demonstrate the scalability and superiority of IBQ, achieving competitive results on both reconstruction ($1.00$ rFID) and autoregressive visual generation ($2.05$ gFID). The code and models are available at https://github.com/TencentARC/SEED-Voken.
Under stringent privacy constraints, whether federated recommendation systems can achieve group fairness remains an inadequately explored question. Taking gender fairness as a representative issue, we identify three phenomena in federated recommendation systems: performance difference, data imbalance, and preference disparity. We discover that the state-of-the-art methods only focus on the first phenomenon. Consequently, their imposition of inappropriate fairness constraints detrimentally affects the model training. Moreover, due to insufficient sensitive attribute protection of existing works, we can infer the gender of all users with 99.90% accuracy even with the addition of maximal noise. In this work, we propose Privacy-Preserving Orthogonal Aggregation (PPOA), which employs the secure aggregation scheme and quantization technique, to prevent the suppression of minority groups by the majority and preserve the distinct preferences for better group fairness. PPOA can assist different groups in obtaining their respective model aggregation results through a designed orthogonal mapping while keeping their attributes private. Experimental results on three real-world datasets demonstrate that PPOA enhances recommendation effectiveness for both females and males by up to 8.25% and 6.36%, respectively, with a maximum overall improvement of 7.30%, and achieves optimal fairness in most cases. Extensive ablation experiments and visualizations indicate that PPOA successfully maintains preferences for different gender groups.
Recent advancements in fine-tuning proprietary language models enable customized applications across various domains but also introduce two major challenges: high resource demands and security risks. Regarding resource demands, recent work proposes novel partial compression, such as BitDelta, to quantize the delta weights between the fine-tuned model and base model. Regarding the security risks, user-defined fine-tuning can introduce security vulnerabilities, such as alignment issues, backdoor attacks, and hallucinations. However, most of the current efforts in security assessment focus on the full-precision or full-compression models, it is not well-discussed how the partial compression methods affect security concerns. To bridge this gap, we evaluate the robustness of delta-weight quantization against these security threats. In this paper, we uncover a "free lunch" phenomenon: partial compression can enhance model security against fine-tuning-based attacks with bearable utility loss. Using Llama-2-7b-chat as a case study, we show that, with under 10% utility degradation, the partial compression mitigates alignment-breaking risks by up to 66.17%, harmful backdoor vulnerabilities by 64.46%, and targeted output manipulation risks by up to 90.53%. We further apply LogitLens to visualize internal state transformations during forward passes, suggesting mechanisms for both security failure and recovery in standard versus compressed fine-tuning. This work offers new insights into selecting effective delta compression methods for secure, resource-efficient multi-tenant services.
The tokenization of speech with neural audio codec models is a vital part of modern AI pipelines for the generation or understanding of speech, alone or in a multimodal context. Traditionally such tokenization models have concentrated on low parameter-count architectures using only components with strong inductive biases. In this work we show that by scaling a transformer architecture with large parameter count to this problem, and applying a flexible Finite Scalar Quantization (FSQ) based bottleneck, it is possible to reach state-of-the-art speech quality at extremely low bit-rates of $400$ or $700$ bits-per-second. The trained models strongly out-perform existing baselines in both objective and subjective tests.
The advancement of large Vision-Language-Action (VLA) models has significantly improved robotic manipulation in terms of language-guided task execution and generalization to unseen scenarios. While existing VLAs adapted from pretrained large Vision-Language-Models (VLM) have demonstrated promising generalizability, their task performance is still unsatisfactory as indicated by the low tasks success rates in different environments. In this paper, we present a new advanced VLA architecture derived from VLM. Unlike previous works that directly repurpose VLM for action prediction by simple action quantization, we propose a omponentized VLA architecture that has a specialized action module conditioned on VLM output. We systematically study the design of the action module and demonstrates the strong performance enhancement with diffusion action transformers for action sequence modeling, as well as their favorable scaling behaviors. We also conduct comprehensive experiments and ablation studies to evaluate the efficacy of our models with varied designs. The evaluation on 5 robot embodiments in simulation and real work shows that our model not only significantly surpasses existing VLAs in task performance and but also exhibits remarkable adaptation to new robots and generalization to unseen objects and backgrounds. It exceeds the average success rates of OpenVLA which has similar model size (7B) with ours by over 35% in simulated evaluation and 55% in real robot experiments. It also outperforms the large RT-2-X model (55B) by 18% absolute success rates in simulation. Code and models can be found on our project page (https://cogact.github.io/).
Human motion is inherently continuous and dynamic, posing significant challenges for generative models. While discrete generation methods are widely used, they suffer from limited expressiveness and frame-wise noise artifacts. In contrast, continuous approaches produce smoother, more natural motion but often struggle to adhere to conditioning signals due to high-dimensional complexity and limited training data. To resolve this discord between discrete and continuous representations, we introduce DisCoRD: Discrete Tokens to Continuous Motion via Rectified Flow Decoding, a novel method that leverages rectified flow to decode discrete motion tokens in the continuous, raw motion space. Our core idea is to frame token decoding as a conditional generation task, ensuring that DisCoRD captures fine-grained dynamics and achieves smoother, more natural motions. Compatible with any discrete-based framework, our method enhances naturalness without compromising faithfulness to the conditioning signals on diverse settings. Extensive evaluations Our project page is available at: https://whwjdqls.github.io/discord.github.io/.
Sparse mixture of experts (SMoE) is an effective solution for scaling up model capacity without increasing the computational costs. A crucial component of SMoE is the router, responsible for directing the input to relevant experts; however, it also presents a major weakness, leading to routing inconsistencies and representation collapse issues. Instead of fixing the router like previous works, we propose an alternative that assigns experts to input via indirection, which employs the discrete representation of input that points to the expert. The discrete representations are learnt via vector quantization, resulting in a new architecture dubbed Vector-Quantized Mixture of Experts (VQMoE). We provide theoretical support and empirical evidence demonstrating the VQMoE's ability to overcome the challenges present in traditional routers. Through extensive evaluations on both large language models and vision tasks for pre-training and fine-tuning, we show that VQMoE achieves a 28% improvement in robustness compared to other SMoE routing methods, while maintaining strong performance in fine-tuning tasks.
A widely-used technique in designing energy-efficient deep neural network (DNN) accelerators is quantization. Recent progress in this direction has reduced the bitwidths used in DNN down to 2. Meanwhile, many prior works apply approximate multipliers (AppMuls) in designing DNN accelerators to lower their energy consumption. Unfortunately, these works still assume a bitwidth much larger than 2, which falls far behind the state-of-the-art in quantization area and even challenges the meaningfulness of applying AppMuls in DNN accelerators, since a high-bitwidth AppMul consumes much more energy than a low-bitwidth exact multiplier! Thus, an important problem to study is: Can approximate multipliers be effectively applied to quantized DNN models with very low bitwidths? In this work, we give an affirmative answer to this question and present a systematic solution that achieves the answer: FAMES, a fast approximate multiplier substitution method for mixed-precision DNNs. Our experiments demonstrate an average 28.67% energy reduction on state-of-the-art mixed-precision quantized models with bitwidths as low as 2 bits and accuracy losses kept under 1%. Additionally, our approach is up to 300x faster than previous genetic algorithm-based methods.
Recent research efforts focus on reducing the computational and memory overheads of Large Language Models (LLMs) to make them feasible on resource-constrained devices. Despite advancements in compression techniques, non-linear operators like Softmax and Layernorm remain bottlenecks due to their sensitivity to quantization. We propose SoftmAP, a software-hardware co-design methodology that implements an integer-only low-precision Softmax using In-Memory Compute (IMC) hardware. Our method achieves up to three orders of magnitude improvement in the energy-delay product compared to A100 and RTX3090 GPUs, making LLMs more deployable without compromising performance.
Quantizing large language models has become a standard way to reduce their memory and computational costs. Typically, existing methods focus on breaking down the problem into individual layer-wise sub-problems, and minimizing per-layer error, measured via various metrics. Yet, this approach currently lacks theoretical justification and the metrics employed may be sub-optimal. In this paper, we present a "linearity theorem" establishing a direct relationship between the layer-wise $\ell_2$ reconstruction error and the model perplexity increase due to quantization. This insight enables two novel applications: (1) a simple data-free LLM quantization method using Hadamard rotations and MSE-optimal grids, dubbed HIGGS, which outperforms all prior data-free approaches such as the extremely popular NF4 quantized format, and (2) an optimal solution to the problem of finding non-uniform per-layer quantization levels which match a given compression constraint in the medium-bitwidth regime, obtained by reduction to dynamic programming. On the practical side, we demonstrate improved accuracy-compression trade-offs on Llama-3.1 and 3.2-family models, as well as on Qwen-family models. Further, we show that our method can be efficiently supported in terms of GPU kernels at various batch sizes, advancing both data-free and non-uniform quantization for LLMs.
Harnessing the local topography of the loss landscape is a central challenge in advanced optimization tasks. By accounting for the effect of potential parameter changes, we can alter the model more efficiently. Contrary to standard assumptions, we find that the Hessian does not always approximate loss curvature well, particularly near gradient discontinuities, which commonly arise in deep learning architectures. We present a new conceptual framework to understand how curvature of expected changes in loss emerges in architectures with many rectified linear units. Each ReLU creates a parameter boundary that, when crossed, induces a pseudorandom gradient perturbation. Our derivations show how these discontinuities combine to form a glass-like structure, similar to amorphous solids that contain microscopic domains of strong, but random, atomic alignment. By estimating the density of the resulting gradient variations, we can bound how the loss may change with parameter movement. Our analysis includes the optimal kernel and sample distribution for approximating glass density from ordinary gradient evaluations. We also derive the optimal modification to quasi-Newton steps that incorporate both glass and Hessian terms, as well as certain exactness properties that are possible with Nesterov-accelerated gradient updates. Our algorithm, Alice, tests these techniques to determine which curvature terms are most impactful for training a given architecture and dataset. Additional safeguards enforce stable exploitation through step bounds that expand on the functionality of Adam. These theoretical and experimental tools lay groundwork to improve future efforts (e.g., pruning and quantization) by providing new insight into the loss landscape.
We reveal that low-bit quantization favors undertrained large language models (LLMs) by observing that models with larger sizes or fewer training tokens experience less quantization-induced degradation (QiD) when applying low-bit quantization, whereas smaller models with extensive training tokens suffer significant QiD. To gain deeper insights into this trend, we study over 1500 quantized LLM checkpoints of various sizes and at different training levels (undertrained or fully trained) in a controlled setting, deriving scaling laws for understanding the relationship between QiD and factors such as the number of training tokens, model size and bit width. With the derived scaling laws, we propose a novel perspective that we can use QiD to measure an LLM's training levels and determine the number of training tokens required for fully training LLMs of various sizes. Moreover, we use the scaling laws to predict the quantization performance of different-sized LLMs trained with 100 trillion tokens. Our projection shows that the low-bit quantization performance of future models, which are expected to be trained with over 100 trillion tokens, may NOT be desirable. This poses a potential challenge for low-bit quantization in the future and highlights the need for awareness of a model's training level when evaluating low-bit quantization research. To facilitate future research on this problem, we release all the 1500+ quantized checkpoints used in this work at https://huggingface.co/Xu-Ouyang.
Visual Autoregressive (VAR) has emerged as a promising approach in image generation, offering competitive potential and performance comparable to diffusion-based models. However, current AR-based visual generation models require substantial computational resources, limiting their applicability on resource-constrained devices. To address this issue, we conducted analysis and identified significant redundancy in three dimensions of the VAR model: (1) the attention map, (2) the attention outputs when using classifier free guidance, and (3) the data precision. Correspondingly, we proposed efficient attention mechanism and low-bit quantization method to enhance the efficiency of VAR models while maintaining performance. With negligible performance lost (less than 0.056 FID increase), we could achieve 85.2% reduction in attention computation, 50% reduction in overall memory and 1.5x latency reduction. To ensure deployment feasibility, we developed efficient training-free compression techniques and analyze the deployment feasibility and efficiency gain of each technique.
Diffusion-based image super-resolution (SR) models have shown superior performance at the cost of multiple denoising steps. However, even though the denoising step has been reduced to one, they require high computational costs and storage requirements, making it difficult for deployment on hardware devices. To address these issues, we propose a novel post-training quantization approach with adaptive scale in one-step diffusion (OSD) image SR, PassionSR. First, we simplify OSD model to two core components, UNet and Variational Autoencoder (VAE) by removing the CLIPEncoder. Secondly, we propose Learnable Boundary Quantizer (LBQ) and Learnable Equivalent Transformation (LET) to optimize the quantization process and manipulate activation distributions for better quantization. Finally, we design a Distributed Quantization Calibration (DQC) strategy that stabilizes the training of quantized parameters for rapid convergence. Comprehensive experiments demonstrate that PassionSR with 8-bit and 6-bit obtains comparable visual results with full-precision model. Moreover, our PassionSR achieves significant advantages over recent leading low-bit quantization methods for image SR. Our code will be at https://github.com/libozhu03/PassionSR.