Large language models (LLMs) exhibit excellent performance in various tasks. However, the memory requirements of LLMs present a great challenge when deploying on memory-limited devices, even for quantized LLMs. This paper introduces a framework to compress LLM after quantization further, achieving about 2.2x compression ratio. A compression-aware quantization is first proposed to enhance model weight compressibility by re-scaling the model parameters before quantization, followed by a pruning method to improve further. Upon this, we notice that decompression can be a bottleneck during practical scenarios. We then give a detailed analysis of the trade-off between memory usage and latency brought by the proposed method. A speed-adaptive method is proposed to overcome it. The experimental results show inference with the compressed model can achieve a 40% reduction in memory size with negligible loss in accuracy and inference speed.
For the efficient inference of Large Language Models (LLMs), the effective compression of key-value (KV) cache is essential. Three main types of KV cache compression techniques, namely sparsity, channel compression, and quantization, have been identified. This study presents SVDq, a Singular Value Decomposition (SVD) - based mixed precision quantization method for K cache. Initially, K cache is transformed into latent channels using SVD basis representations. Since the values in latent channels decay rapidly and become negligible after only a few latent channels, our method then incorporates importance-aware quantization and compression for latent channels. This enables the effective allocation of higher precision to more significant channels. Theoretically, we prove that SVDq results in quantization errors (x0.1 or even lower) that are much lower than those of per-channel key quantization in the original space. Our findings based on RULER and LongBench benchmarks demonstrate that SVDq can achieve an equivalent key cache precision as low as 1.25-bit. When combined with key sparsity, it can reach a key compression ratio of up to 410x for attention computation, all while maintaining comparable model performance. Notably, our method is nearly lossless for LongBench datasets. This indicates that SVDq enables high-precision low-bit quantization, providing a more efficient solution for KV cache compression in LLMs.
We introduce Dr. Splat, a novel approach for open-vocabulary 3D scene understanding leveraging 3D Gaussian Splatting. Unlike existing language-embedded 3DGS methods, which rely on a rendering process, our method directly associates language-aligned CLIP embeddings with 3D Gaussians for holistic 3D scene understanding. The key of our method is a language feature registration technique where CLIP embeddings are assigned to the dominant Gaussians intersected by each pixel-ray. Moreover, we integrate Product Quantization (PQ) trained on general large-scale image data to compactly represent embeddings without per-scene optimization. Experiments demonstrate that our approach significantly outperforms existing approaches in 3D perception benchmarks, such as open-vocabulary 3D semantic segmentation, 3D object localization, and 3D object selection tasks. For video results, please visit : https://drsplat.github.io/
We consider the problem of exact recovery of a $k$-sparse binary vector from generalized linear measurements (such as logistic regression). We analyze the linear estimation algorithm (Plan, Vershynin, Yudovina, 2017), and also show information theoretic lower bounds on the number of required measurements. As a consequence of our results, for noisy one bit quantized linear measurements ($\mathsf{1bCSbinary}$), we obtain a sample complexity of $O((k+\sigma^2)\log{n})$, where $\sigma^2$ is the noise variance. This is shown to be optimal due to the information theoretic lower bound. We also obtain tight sample complexity characterization for logistic regression. Since $\mathsf{1bCSbinary}$ is a strictly harder problem than noisy linear measurements ($\mathsf{SparseLinearReg}$) because of added quantization, the same sample complexity is achievable for $\mathsf{SparseLinearReg}$. While this sample complexity can be obtained via the popular lasso algorithm, linear estimation is computationally more efficient. Our lower bound holds for any set of measurements for $\mathsf{SparseLinearReg}$, (similar bound was known for Gaussian measurement matrices) and is closely matched by the maximum-likelihood upper bound. For $\mathsf{SparseLinearReg}$, it was conjectured in Gamarnik and Zadik, 2017 that there is a statistical-computational gap and the number of measurements should be at least $(2k+\sigma^2)\log{n}$ for efficient algorithms to exist. It is worth noting that our results imply that there is no such statistical-computational gap for $\mathsf{1bCSbinary}$ and logistic regression.
Quantizing model weights is critical for reducing the communication and inference costs of large models. However, quantizing models -- especially to low precisions like int4 or int2 -- requires a trade-off in model quality; int2, in particular, is known to severely degrade model quality. Consequently, practitioners are often forced to maintain multiple models with different quantization levels or serve a single model that best satisfies the quality-latency trade-off. On the other hand, integer data types, such as int8, inherently possess a nested (Matryoshka) structure where smaller bit-width integers, like int4 or int2, are nested within the most significant bits. Leveraging this insight, in this paper, we propose Matryoshka Quantization (\alg), a novel multi-scale quantization technique that alleviates the aforementioned challenge. This technique allows us to train and maintain a single quantized model but serve it with the precision demanded by the deployment. Furthermore, leveraging \alg's co-training and co-distillation regularization, int2 precision models extracted by \alg outperform standard int2 quantization by up to to 4\% and 7\% with OmniQuant and QAT as base algorithms respectively. Finally, we demonstrate that by using an extra bit to represent outliers, a model with an effective precision of 2.05-bit gives an additional 6\% improvement with OmniQuant as the base algorithm.
Structured pruning and quantization are fundamental techniques used to reduce the size of deep neural networks (DNNs) and typically are applied independently. Applying these techniques jointly via co-optimization has the potential to produce smaller, high-quality models. However, existing joint schemes are not widely used because of (1) engineering difficulties (complicated multi-stage processes), (2) black-box optimization (extensive hyperparameter tuning to control the overall compression), and (3) insufficient architecture generalization. To address these limitations, we present the framework GETA, which automatically and efficiently performs joint structured pruning and quantization-aware training on any DNNs. GETA introduces three key innovations: (i) a quantization-aware dependency graph (QADG) that constructs a pruning search space for generic quantization-aware DNN, (ii) a partially projected stochastic gradient method that guarantees layerwise bit constraints are satisfied, and (iii) a new joint learning strategy that incorporates interpretable relationships between pruning and quantization. We present numerical experiments on both convolutional neural networks and transformer architectures that show that our approach achieves competitive (often superior) performance compared to existing joint pruning and quantization methods.
The increasing computational demands of transformer models in time series classification necessitate effective optimization strategies for energy-efficient deployment. Our study presents a systematic investigation of optimization techniques, focusing on structured pruning and quantization methods for transformer architectures. Through extensive experimentation on three distinct datasets (RefrigerationDevices, ElectricDevices, and PLAID), we quantitatively evaluate model performance and energy efficiency across different transformer configurations. Our experimental results demonstrate that static quantization reduces energy consumption by 29.14% while maintaining classification performance, and L1 pruning achieves a 63% improvement in inference speed with minimal accuracy degradation. Our findings provide valuable insights into the effectiveness of optimization strategies for transformer-based time series classification, establishing a foundation for efficient model deployment in resource-constrained environments.
In the rapidly evolving landscape of neural network security, the resilience of neural networks against bit-flip attacks (i.e., an attacker maliciously flips an extremely small amount of bits within its parameter storage memory system to induce harmful behavior), has emerged as a relevant area of research. Existing studies suggest that quantization may serve as a viable defense against such attacks. Recognizing the documented susceptibility of real-valued neural networks to such attacks and the comparative robustness of quantized neural networks (QNNs), in this work, we introduce BFAVerifier, the first verification framework designed to formally verify the absence of bit-flip attacks or to identify all vulnerable parameters in a sound and rigorous manner. BFAVerifier comprises two integral components: an abstraction-based method and an MILP-based method. Specifically, we first conduct a reachability analysis with respect to symbolic parameters that represent the potential bit-flip attacks, based on a novel abstract domain with a sound guarantee. If the reachability analysis fails to prove the resilience of such attacks, then we encode this verification problem into an equivalent MILP problem which can be solved by off-the-shelf solvers. Therefore, BFAVerifier is sound, complete, and reasonably efficient. We conduct extensive experiments, which demonstrate its effectiveness and efficiency across various network architectures, quantization bit-widths, and adversary capabilities.
Recent advancements in Neural Audio Codec (NAC) models have inspired their use in various speech processing tasks, including speech enhancement (SE). In this work, we propose a novel, efficient SE approach by leveraging the pre-quantization output of a pretrained NAC encoder. Unlike prior NAC-based SE methods, which process discrete speech tokens using Language Models (LMs), we perform SE within the continuous embedding space of the pretrained NAC, which is highly compressed along the time dimension for efficient representation. Our lightweight SE model, optimized through an embedding-level loss, delivers results comparable to SE baselines trained on larger datasets, with a significantly lower real-time factor of 0.005. Additionally, our method achieves a low GMAC of 3.94, reducing complexity 18-fold compared to Sepformer in a simulated cloud-based audio transmission environment. This work highlights a new, efficient NAC-based SE solution, particularly suitable for cloud applications where NAC is used to compress audio before transmission. Copyright 20XX IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
The exponential growth in parameter size and computational complexity of deep models poses significant challenges for efficient deployment. The core problem of existing compression methods is that different layers of the model have significant differences in their tolerance to compression levels. For instance, the first layer of a model can typically sustain a higher compression level compared to the last layer without compromising performance. Thus, the key challenge lies in how to allocate compression levels across layers in a way that minimizes performance loss while maximizing parameter reduction. To address this challenge, we propose a Compression Error Theory (CET) framework, designed to determine the optimal compression level for each layer. Taking quantization as an example, CET leverages differential expansion and algebraic geometry to reconstruct the quadratic form of quantization error as ellipsoids and hyperbolic paraboloids, and utilizes their geometric structures to define an error subspace. To identify the error subspace with minimal performance loss, by performing orthogonal decomposition of the geometric space, CET transforms the optimization process of the error subspace into a complementary problem. The final theoretical analysis shows that constructing the quantization subspace along the major axis results in minimal performance degradation. Through experimental verification of the theory, CET can greatly retain performance while compressing. Specifically, on the ResNet-34 model, CET achieves nearly 11$\times$ parameter compression while even surpassing performance comparable to the original model.
Large Language Models (LLMs) have emerged as powerful tools for addressing modern challenges and enabling practical applications. However, their computational expense remains a significant barrier to widespread adoption. Quantization has emerged as a promising technique to democratize access and enable low resource device deployment. Despite these advancements, the safety and trustworthiness of quantized models remain underexplored, as prior studies often overlook contemporary architectures and rely on overly simplistic benchmarks and evaluations. To address this gap, we introduce OpenSafetyMini, a novel open-ended safety dataset designed to better distinguish between models. We evaluate 4 state-of-the-art quantization techniques across LLaMA and Mistral models using 4 benchmarks, including human evaluations. Our findings reveal that the optimal quantization method varies for 4-bit precision, while vector quantization techniques deliver the best safety and trustworthiness performance at 2-bit precision, providing foundation for future research.
We propose Rotate, Clip, and Partition (RCP), a quantization-aware training (QAT) approach that first realizes extreme compression of LLMs with W2A4KV4(2-bit weight, 4-bit activation, and 4-bit KV cache) configuration. RCP integrates recent rotation techniques with a novel non-uniform weight quantizer design, by quantitatively analyzing the impact of random rotation on 2-bit weight quantization. Our weight quantizer features Learnable Direct Partitioning (LDP), which introduces learnable parameters to directly learn non-uniform intervals jointly with LLM weights. We also present a specialized GPU kernel that supports GEMV on non-uniform W2A4. Experiments show that RCP can compress LLaMA-2-7B to W2A4KV4 with a loss of only 2.84 WikiText2 ppl and 5.29 times reduced memory footprint. Furthermore, RCP can quantize challenging mobile-targeted LLaMA-3.2 models and domain-specific WizardCoder-7B and MetaMath-7B with no critical problems such as convergence failure and repetition. Code will be made available at blind_review.
Advancements in embedded systems and Artificial Intelligence (AI) have enhanced the capabilities of Unmanned Aircraft Vehicles (UAVs) in computer vision. However, the integration of AI techniques o-nboard drones is constrained by their processing capabilities. In this sense, this study evaluates the deployment of object detection models (YOLOv8n and YOLOv8s) on both resource-constrained edge devices and cloud environments. The objective is to carry out a comparative performance analysis using a representative real-time UAV image processing pipeline. Specifically, the NVIDIA Jetson Orin Nano, Orin NX, and Raspberry Pi 5 (RPI5) devices have been tested to measure their detection accuracy, inference speed, and energy consumption, and the effects of post-training quantization (PTQ). The results show that YOLOv8n surpasses YOLOv8s in its inference speed, achieving 52 FPS on the Jetson Orin NX and 65 fps with INT8 quantization. Conversely, the RPI5 failed to satisfy the real-time processing needs in spite of its suitability for low-energy consumption applications. An analysis of both the cloud-based and edge-based end-to-end processing times showed that increased communication latencies hindered real-time applications, revealing trade-offs between edge (low latency) and cloud processing (quick processing). Overall, these findings contribute to providing recommendations and optimization strategies for the deployment of AI models on UAVs.
In recent years, learned image compression (LIC) methods have achieved significant performance improvements. However, obtaining a more compact latent representation and reducing the impact of quantization errors remain key challenges in the field of LIC. To address these challenges, we propose a feature extraction module, a feature refinement module, and a feature enhancement module. Our feature extraction module shuffles the pixels in the image, splits the resulting image into sub-images, and extracts coarse features from the sub-images. Our feature refinement module stacks the coarse features and uses an attention refinement block composed of concatenated three-dimensional convolution residual blocks to learn more compact latent features by exploiting correlations across channels, within sub-images (intra-sub-image correlations), and across sub-images (inter-sub-image correlations). Our feature enhancement module reduces information loss in the decoded features following quantization. We also propose a quantization error compensation module that mitigates the quantization mismatch between training and testing. Our four modules can be readily integrated into state-of-the-art LIC methods. Experiments show that combining our modules with Tiny-LIC outperforms existing LIC methods and image compression standards in terms of peak signal-to-noise ratio (PSNR) and multi-scale structural similarity (MS-SSIM) on the Kodak dataset and the CLIC dataset.
The learned image compression (LIC) methods have already surpassed traditional techniques in compressing natural scene (NS) images. However, directly applying these methods to screen content (SC) images, which possess distinct characteristics such as sharp edges, repetitive patterns, embedded text and graphics, yields suboptimal results. This paper addresses three key challenges in SC image compression: learning compact latent features, adapting quantization step sizes, and the lack of large SC datasets. To overcome these challenges, we propose a novel compression method that employs a multi-frequency two-stage octave residual block (MToRB) for feature extraction, a cascaded triple-scale feature fusion residual block (CTSFRB) for multi-scale feature integration and a multi-frequency context interaction module (MFCIM) to reduce inter-frequency correlations. Additionally, we introduce an adaptive quantization module that learns scaled uniform noise for each frequency component, enabling flexible control over quantization granularity. Furthermore, we construct a large SC image compression dataset (SDU-SCICD10K), which includes over 10,000 images spanning basic SC images, computer-rendered images, and mixed NS and SC images from both PC and mobile platforms. Experimental results demonstrate that our approach significantly improves SC image compression performance, outperforming traditional standards and state-of-the-art learning-based methods in terms of peak signal-to-noise ratio (PSNR) and multi-scale structural similarity (MS-SSIM).
This paper introduces an information-aware quantization framework that adaptively compresses the key-value (KV) cache in large language models (LLMs). Although prior work has underscored the distinct roles of key and value cache during inference, our systematic analysis -- examining singular value distributions, spectral norms, and Frobenius norms -- reveals, for the first time, that key matrices consistently exhibit higher norm values and are more sensitive to quantization than value matrices. Furthermore, our theoretical analysis shows that matrices with higher spectral norms amplify quantization errors more significantly. Motivated by these insights, we propose a mixed-precision quantization strategy, KV-AdaQuant, which allocates more bit-width for keys and fewer for values since key matrices have higher norm values. With the same total KV bit budget, this approach effectively mitigates error propagation across transformer layers while achieving significant memory savings. Our extensive experiments on multiple LLMs (1B--70B) demonstrate that our mixed-precision quantization scheme maintains high model accuracy even under aggressive compression. For instance, using 4-bit for Key and 2-bit for Value achieves an accuracy of 75.2%, whereas reversing the assignment (2-bit for Key and 4-bit for Value) yields only 54.7% accuracy. The code is available at https://tinyurl.com/kv-adaquant
Large language models (LLMs) have demonstrated remarkable performance across a wide range of industrial applications, from search and recommendations to generative tasks. Although scaling laws indicate that larger models generally yield better generalization and performance, their substantial computational requirements often render them impractical for many real-world scenarios at scale. In this paper, we present methods and insights for training small language models (SLMs) that deliver high performance and efficiency in deployment. We focus on two key techniques: (1) knowledge distillation and (2) model compression via quantization and pruning. These approaches enable SLMs to retain much of the quality of their larger counterparts while significantly reducing training, serving costs, and latency. We detail the impact of these techniques on a variety of use cases at a large professional social network platform and share deployment lessons - including hardware optimization strategies that enhance speed and throughput for both predictive and reasoning-based applications.
Multi-head Latent Attention (MLA) is an innovative architecture proposed by DeepSeek, designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector. Compared to MLA, standard LLMs employing Multi-Head Attention (MHA) and its variants such as Grouped-Query Attention (GQA) exhibit significant cost disadvantages. Enabling well-trained LLMs (e.g., Llama) to rapidly adapt to MLA without pre-training from scratch is both meaningful and challenging. This paper proposes the first data-efficient fine-tuning method for transitioning from MHA to MLA (MHA2MLA), which includes two key components: for partial-RoPE, we remove RoPE from dimensions of queries and keys that contribute less to the attention scores, for low-rank approximation, we introduce joint SVD approximations based on the pre-trained parameters of keys and values. These carefully designed strategies enable MHA2MLA to recover performance using only a small fraction (0.3% to 0.6%) of the data, significantly reducing inference costs while seamlessly integrating with compression techniques such as KV cache quantization. For example, the KV cache size of Llama2-7B is reduced by 92.19%, with only a 0.5% drop in LongBench performance.
Long context large language models (LLMs) pose significant challenges for efficient serving due to the large memory footprint and high access overhead of KV cache. Retrieval-based KV cache reduction methods can mitigate these challenges, typically by offloading the complete KV cache to CPU and retrieving necessary tokens on demand during inference. However, these methods still suffer from unsatisfactory accuracy degradation and extra retrieval overhead. To address these limitations, this paper proposes A$^2$ATS, a novel retrieval-based KV cache reduction method. A$^2$ATS aims to obtain an accurate approximation of attention scores by applying the vector quantization technique to key states, thereby enabling efficient and precise retrieval of the top-K tokens. First, we propose Windowed Rotary Position Embedding, which decouples the positional dependency from query and key states after position embedding. Then, we propose query-aware vector quantization that optimizes the objective of attention score approximation directly. Finally, we design the heterogeneous inference architecture for KV cache offloading, enabling long context serving with larger batch sizes. Experimental results demonstrate that A$^2$ATS can achieve a lower performance degradation with similar or lower overhead compared to existing methods, thereby increasing long context serving throughput by up to $2.7 \times$.
Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy without additional GPU overhead. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99\%. Additionally, Fate integrates tailored quantization strategies for cache optimization and IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.