State-Space Models (SSMs) have attracted considerable attention in Image Restoration (IR) due to their ability to scale linearly sequence length while effectively capturing long-distance dependencies. However, deploying SSMs to edge devices is challenging due to the constraints in memory, computing capacity, and power consumption, underscoring the need for efficient compression strategies. While low-bit quantization is an efficient model compression strategy for reducing size and accelerating IR tasks, SSM suffers substantial performance drops at ultra-low bit-widths (2-4 bits), primarily due to outliers that exacerbate quantization error. To address this challenge, we propose Q-MambaIR, an accurate, efficient, and flexible Quantized Mamba for IR tasks. Specifically, we introduce a Statistical Dynamic-balancing Learnable Scalar (DLS) to dynamically adjust the quantization mapping range, thereby mitigating the peak truncation loss caused by extreme values. Furthermore, we design a Range-floating Flexible Allocator (RFA) with an adaptive threshold to flexibly round values. This approach preserves high-frequency details and maintains the SSM's feature extraction capability. Notably, RFA also enables pre-deployment weight quantization, striking a balance between computational efficiency and model accuracy. Extensive experiments on IR tasks demonstrate that Q-MambaIR consistently outperforms existing quantized SSMs, achieving much higher state-of-the-art (SOTA) accuracy results with only a negligible increase in training computation and storage saving.
Large language models (LLMs) hold great promise for specialized scientific domains such as materials science, yet adapting them efficiently and accurately to domain-specific knowledge remains challenging due to limited data and high knowledge density. We propose a two-stage framework that combines structured model compression with a scientific fine-tuning regimen to address this challenge. In the compression stage, we decompose the LLM's weight matrices into local low-rank "rank blocks" and arrange these blocks in a Penrose-like non-periodic tiling pattern. Each block is then compacted via spectral transformations (e.g., discrete cosine or Fourier transforms), and a Kullback-Leibler (KL) divergence-based alignment loss preserves the distributional similarity between the compressed model's representations and those of the original full model. In the adaptation stage, the compressed model is further tuned using a human-like scientific reading protocol: it processes technical materials science documents section by section, engaging in a structured question-and-answer routine for each section. This section-wise Q&A fine-tuning strategy extracts explicit reasoning traces and gradually injects domain knowledge, while minimizing catastrophic forgetting of the model's general language capabilities. By balancing efficient compression with targeted adaptation, our two-stage approach enables precise specialization of LLMs to high-value domains under data-scarce conditions. We present this principled yet exploratory pipeline and outline its potential for advancing materials science knowledge integration, laying the groundwork for comprehensive empirical evaluation in future work.
Artificial intelligence (AI) promises to revolutionize the design, optimization and management of next-generation communication systems. In this article, we explore the integration of large AI models (LAMs) into semantic communications (SemCom) by leveraging their multi-modal data processing and generation capabilities. Although LAMs bring unprecedented abilities to extract semantics from raw data, this integration entails multifaceted challenges including high resource demands, model complexity, and the need for adaptability across diverse modalities and tasks. To overcome these challenges, we propose a LAM-based multi-task SemCom (MTSC) architecture, which includes an adaptive model compression strategy and a federated split fine-tuning approach to facilitate the efficient deployment of LAM-based semantic models in resource-limited networks. Furthermore, a retrieval-augmented generation scheme is implemented to synthesize the most recent local and global knowledge bases to enhance the accuracy of semantic extraction and content generation, thereby improving the inference performance. Finally, simulation results demonstrate the efficacy of the proposed LAM-based MTSC architecture, highlighting the performance enhancements across various downstream tasks under varying channel conditions.
Linear recurrent neural networks enable powerful long-range sequence modeling with constant memory usage and time-per-token during inference. These architectures hold promise for streaming applications at the edge, but deployment in resource-constrained environments requires hardware-aware optimizations to minimize latency and energy consumption. Unstructured sparsity offers a compelling solution, enabling substantial reductions in compute and memory requirements--when accelerated by compatible hardware platforms. In this paper, we conduct a scaling study to investigate the Pareto front of performance and efficiency across inference compute budgets. We find that highly sparse linear RNNs consistently achieve better efficiency-performance trade-offs than dense baselines, with 2x less compute and 36% less memory at iso-accuracy. Our models achieve state-of-the-art results on a real-time streaming task for audio denoising. By quantizing our sparse models to fixed-point arithmetic and deploying them on the Intel Loihi 2 neuromorphic chip for real-time processing, we translate model compression into tangible gains of 42x lower latency and 149x lower energy consumption compared to a dense model on an edge GPU. Our findings showcase the transformative potential of unstructured sparsity, paving the way for highly efficient recurrent neural networks in real-world, resource-constrained environments.
With the advances in artificial intelligence, Mix-of-Experts (MoE) has become the main form of Large Language Models (LLMs), and its demand for model compression is increasing. Quantization is an effective method that not only compresses the models but also significantly accelerates their performance. Existing quantization methods have gradually shifted the focus from parameter scaling to the analysis of data distributions. However, their analysis is designed for dense LLMs and relies on the simple one-model-all-data mapping, which is unsuitable for MoEs. This paper proposes a new quantization framework called MoQa. MoQa decouples the data-model distribution complexity of MoEs in multiple analysis stages, quantitively revealing the dynamics during sparse data activation, data-parameter mapping, and inter-expert correlations. Based on these, MoQa identifies particular experts' and parameters' significance with optimal data-model distribution awareness and proposes a series of fine-grained mix-quantization strategies adaptive to various data activation and expert combination scenarios. Moreover, MoQa discusses the limitations of existing quantization and analyzes the impact of each stage analysis, showing novel insights for MoE quantization. Experiments show that MoQa achieves a 1.69~2.18 perplexity decrease in language modeling tasks and a 1.58%~8.91% accuracy improvement in zero-shot inference tasks. We believe MoQa will play a role in future MoE construction, optimization, and compression.
Knowledge distillation has become a cornerstone technique in deep learning, facilitating the transfer of knowledge from complex models to lightweight counterparts. Traditional distillation approaches focus on transferring knowledge at the instance level, but fail to capture nuanced semantic relationships within the data. In response, this paper introduces a novel methodology, Semantics-based Relation Knowledge Distillation (SeRKD), which reimagines knowledge distillation through a semantics-relation lens among each sample. By leveraging semantic components, \ie, superpixels, SeRKD enables a more comprehensive and context-aware transfer of knowledge, which skillfully integrates superpixel-based semantic extraction with relation-based knowledge distillation for a sophisticated model compression and distillation. Particularly, the proposed method is naturally relevant in the domain of Vision Transformers (ViTs), where visual tokens serve as fundamental units of representation. Experimental evaluations on benchmark datasets demonstrate the superiority of SeRKD over existing methods, underscoring its efficacy in enhancing model performance and generalization capabilities.
Transformer-based speech enhancement models yield impressive results. However, their heterogeneous and complex structure restricts model compression potential, resulting in greater complexity and reduced hardware efficiency. Additionally, these models are not tailored for streaming and low-power applications. Addressing these challenges, this paper proposes a low-power streaming speech enhancement accelerator through model and hardware optimization. The proposed high performance model is optimized for hardware execution with the co-design of model compression and target application, which reduces 93.9\% of model size by the proposed domain-aware and streaming-aware pruning techniques. The required latency is further reduced with batch normalization-based transformers. Additionally, we employed softmax-free attention, complemented by an extra batch normalization, facilitating simpler hardware design. The tailored hardware accommodates these diverse computing patterns by breaking them down into element-wise multiplication and accumulation (MAC). This is achieved through a 1-D processing array, utilizing configurable SRAM addressing, thereby minimizing hardware complexities and simplifying zero skipping. Using the TSMC 40nm CMOS process, the final implementation requires merely 207.8K gates and 53.75KB SRAM. It consumes only 8.08 mW for real-time inference at a 62.5MHz frequency.
In this paper, we tackle the critical challenge of compressing large language models (LLMs) to facilitate their practical deployment and broader adoption. We introduce a novel post-training compression paradigm that focuses on low-rank decomposition of LLM weights. Our analysis identifies two main challenges in this task: the variability in LLM activation distributions and handling unseen activations from different datasets and models. To address these challenges, we propose a nested activation-aware framework (NSVD) for LLMs, a training-free approach designed to enhance the accuracy of low-rank decompositions by managing activation outliers through transforming the weight matrix based on activation distribution and the original weight matrix. This method allows for the absorption of outliers into the transformed weight matrix, improving decomposition accuracy. Our comprehensive evaluation across eight datasets and six models from three distinct LLM families demonstrates the superiority of NSVD over current state-of-the-art methods, especially at medium to large compression ratios or in multilingual and multitask settings.
Temporal action detection (TAD) aims to identify and localize action instances in untrimmed videos, which is essential for various video understanding tasks. However, recent improvements in model performance, driven by larger feature extractors and datasets, have led to increased computational demands. This presents a challenge for applications like autonomous driving and robotics, which rely on limited computational resources. While existing channel pruning methods can compress these models, reducing the number of channels often hinders the parallelization efficiency of GPU, due to the inefficient multiplication between small matrices. Instead of pruning channels, we propose a Progressive Block Drop method that reduces model depth while retaining layer width. In this way, we still use large matrices for computation but reduce the number of multiplications. Our approach iteratively removes redundant blocks in two steps: first, we drop blocks with minimal impact on model performance; and second, we employ a parameter-efficient cross-depth alignment technique, fine-tuning the pruned model to restore model accuracy. Our method achieves a 25% reduction in computational overhead on two TAD benchmarks (THUMOS14 and ActivityNet-1.3) to achieve lossless compression. More critically, we empirically show that our method is orthogonal to channel pruning methods and can be combined with it to yield further efficiency gains.
3D Gaussian Splatting (3DGS) enables rapid differentiable rendering for 3D reconstruction and novel view synthesis, leading to its widespread commercial use. Consequently, copyright protection via watermarking has become critical. However, because 3DGS relies on millions of Gaussians, which require gigabytes of storage, efficient transfer and storage require compression. Existing 3DGS watermarking methods are vulnerable to quantization-based compression, often resulting in the loss of the embedded watermark. To address this challenge, we propose a novel watermarking method that ensures watermark robustness after model compression while maintaining high rendering quality. In detail, we incorporate a quantization distortion layer that simulates compression during training, preserving the watermark under quantization-based compression. Also, we propose a learnable watermark embedding feature that embeds the watermark into the anchor feature, ensuring structural consistency and seamless integration into the 3D scene. Furthermore, we present a frequency-aware anchor growing mechanism to enhance image quality in high-frequency regions by effectively identifying Guassians within these regions. Experimental results confirm that our method preserves the watermark and maintains superior image quality under high compression, validating it as a promising approach for a secure 3DGS model.
As large language models (LLMs) scale, model compression is crucial for edge deployment and accessibility. Weight-only quantization reduces model size but suffers from performance degradation at lower bit widths. Moreover, standard finetuning is incompatible with quantized models, and alternative methods often fall short of full finetuning. In this paper, we propose ClusComp, a simple yet effective compression paradigm that clusters weight matrices into codebooks and finetunes them block-by-block. ClusComp (1) achieves superior performance in 2-4 bit quantization, (2) pushes compression to 1-bit while outperforming ultra-low-bit methods with minimal finetuning, and (3) enables efficient finetuning, even surpassing existing quantization-based approaches and rivaling full FP16 finetuning. Notably, ClusComp supports compression and finetuning of 70B LLMs on a single A6000-48GB GPU.
Despite significant advancements, the practical deployment of Large Language Models (LLMs) is often hampered by their immense sizes, highlighting the need for effective compression techniques. Singular Value Decomposition (SVD) is a promising LLM compression technique. However, existing SVD-based compression methods fall short in reducing truncation losses, leading to less competitive performance in compressed models. In this work, we introduce SVD-LLM V2, a SVD-based LLM compression method that optimizes singular value truncation in SVD compression with two techniques. First, SVD-LLM V2 proposes to use theoretical truncation loss of weight matrices to assign a unique compression ratio to each weight matrix at different layers to accommodate weight redundancy heterogeneity. Second, SVD-LLM V2 proposes loss-optimized weight truncation to ensure that the truncated singular values result in a lower and more stable truncation loss in practice. We evaluate SVD-LLM V2 on ten datasets and five LLMs at various scales. Our results show SVD-LLM V2 outperforms state-of-the-art SVD-based LLM compression methods. Our code is available at https://github.com/AIoT-MLSys-Lab/SVD-LLM
This work explores optimizing transformer-based language models by integrating model compression techniques with inhibitor attention, a novel alternative attention mechanism. Inhibitor attention employs Manhattan distances and ReLU activations instead of the matrix multiplications and softmax activation of the conventional scaled dot-product attention. This shift offers potential computational and energy savings while maintaining model effectiveness. We propose further adjustments to improve the inhibitor mechanism's training efficiency and evaluate its performance on the DistilBERT architecture. Our knowledge distillation experiments indicate that the modified inhibitor transformer model can achieve competitive performance on standard NLP benchmarks, including General Language Understanding Evaluation (GLUE) and sentiment analysis tasks.
The challenge of reducing the size of Large Language Models (LLMs) while maintaining their performance has gained significant attention. However, existing methods, such as model distillation and transfer learning, often fail to achieve high accuracy. To address this limitation, we introduce the Branch-Merge distillation approach, which enhances model compression through two phases: (1) the Branch Phase, where knowledge from a large teacher model is \textit{selectively distilled} into specialized student models via domain-specific supervised fine-tuning (SFT); And (2) the Merge Phase, where these student models are merged to enable cross-domain knowledge transfer and improve generalization. We validate our distillation approach using DeepSeek-R1 as the teacher and DeepSeek-R1-Distill-Qwen-32B as the student. The resulting merged model, TinyR1-32B-Preview, outperforms its counterpart DeepSeek-R1-Distill-Qwen-32B across multiple benchmarks, including Mathematics (+5.5 points), Coding (+4.4 points) and Science (+2.9 points), while achieving near-equal performance to DeepSeek-R1 on AIME 2024. The Branch-Merge distillation approach provides a scalable solution for creating smaller, high-performing LLMs with reduced computational cost and time.
Post-training pruning, as one of the key techniques for compressing large language models, plays a vital role in lightweight model deployment and model sparsity. However, current mainstream pruning methods dependent on the Hessian matrix face significant limitations in both pruning speed and practical effectiveness due to the computationally intensive nature of second-order derivative calculations. This paper presents SwiftPrune, a novel Hessian-free weight pruning method that achieves hardware-efficient model compression through two key innovations: 1) SwiftPrune eliminates the need for computationally intensive Hessian matrix calculations by introducing a contribution-based weight metric, which evaluates the importance of weights without relying on second-order derivatives. 2) we employ the Exponentially Weighted Moving Average (EWMA) technique to bypass weight sorting, enabling the selection of weights that contribute most to LLM accuracy and further reducing time complexity. Our approach is extended to support structured sparsity pruning, facilitating efficient execution on modern hardware accelerators. We validate the SwiftPrune on three LLMs (namely LLaMA2, LLaMA3, and Pythia), demonstrating that it significantly enhances compression performance. The experimental findings reveal that SwiftPrune completes the pruning process within seconds, achieving an average speedup of 12.29x (up to 56.02x) over existing SOTA approaches.
Due to the large number of parameters, the inference phase of Large Language Models (LLMs) is resource-intensive. Unlike traditional model compression, which needs retraining, recent dynamic computation methods show that not all components are required for inference, enabling a training-free pipeline. In this paper, we focus on the dynamic depth of LLM generation. A token-position aware layer skipping framework is proposed to save 1.5x times operations efficiently while maintaining performance. We first observed that tokens predicted later have lower perplexity and thus require less computation. Then, we propose a training-free algorithm called Position-Aware Depth Decay Decoding ($D^3$), which leverages a power-law decay function, $\left\lfloor L \times (\alpha^i) \right\rfloor$, to determine the number of layers to retain when generating token $T_i$. Remarkably, without any retraining, the $D^3$ achieves success across a wide range of generation tasks for the first time. Experiments on large language models (\ie the Llama) with $7 \sim 70$ billion parameters show that $D^3$ can achieve an average 1.5x speedup compared with the full-inference pipeline while maintaining comparable performance with nearly no performance drop ($<1\%$) on the GSM8K and BBH benchmarks.
The rapid rise of Language Models (LMs) has expanded the capabilities of natural language processing, powering applications from text generation to complex decision-making. While state-of-the-art LMs often boast hundreds of billions of parameters and are primarily deployed in data centers, recent trends show a growing focus on compact models-typically under 10 billion parameters-enabled by techniques such as quantization and other model compression techniques. This shift paves the way for LMs on edge devices, offering potential benefits such as enhanced privacy, reduced latency, and improved data sovereignty. However, the inherent complexity of even these smaller models, combined with the limited computing resources of edge hardware, raises critical questions about the practical trade-offs in executing LM inference outside the cloud. To address these challenges, we present a comprehensive evaluation of generative LM inference on representative CPU-based and GPU-accelerated edge devices. Our study measures key performance indicators-including memory usage, inference speed, and energy consumption-across various device configurations. Additionally, we examine throughput-energy trade-offs, cost considerations, and usability, alongside an assessment of qualitative model performance. While quantization helps mitigate memory overhead, it does not fully eliminate resource bottlenecks, especially for larger models. Our findings quantify the memory and energy constraints that must be considered for practical real-world deployments, offering concrete insights into the trade-offs between model size, inference performance, and efficiency. The exploration of LMs at the edge is still in its early stages. We hope this study provides a foundation for future research, guiding the refinement of models, the enhancement of inference efficiency, and the advancement of edge-centric AI systems.
Despite recent efforts in understanding the compression impact on large language models (LLMs) in terms of their downstream task performance and trustworthiness on relatively simpler uni-modal benchmarks (for example, question answering, common sense reasoning), their detailed study on multi-modal Large Vision-Language Models (LVLMs) is yet to be unveiled. Towards mitigating this gap, we present LVLM-Compress-Bench, a framework to first thoroughly study the broad impact of compression on the generative performance of LVLMs with multi-modal input driven tasks. In specific, we consider two major classes of compression for autoregressive models, namely KV cache and weight compression, for the dynamically growing intermediate cache and static weights, respectively. We use four LVLM variants of the popular LLaVA framework to present our analysis via integrating various state-of-the-art KV and weight compression methods including uniform, outlier-reduced, and group quantization for the KV cache and weights. With this framework we demonstrate on ten different multi-modal datasets with different capabilities including recognition, knowledge, language generation, spatial awareness, visual reasoning, hallucination and visual illusion identification, toxicity, stereotypes and bias. In specific, our framework demonstrates the compression impact on both general and ethically critical metrics leveraging a combination of real world and synthetic datasets to encompass diverse societal intersectional attributes. Extensive experimental evaluations yield diverse and intriguing observations on the behavior of LVLMs at different quantization budget of KV and weights, in both maintaining and losing performance as compared to the baseline model with FP16 data format. Code will be open-sourced at https://github.com/opengear-project/LVLM-compress-bench.
Dense visual prediction tasks, such as detection and segmentation, are crucial for time-critical applications (e.g., autonomous driving and video surveillance). While deep models achieve strong performance, their efficiency remains a challenge. Knowledge distillation (KD) is an effective model compression technique, but existing feature-based KD methods rely on static, teacher-driven feature selection, failing to adapt to the student's evolving learning state or leverage dynamic student-teacher interactions. To address these limitations, we propose Adaptive student-teacher Cooperative Attention Masking for Knowledge Distillation (ACAM-KD), which introduces two key components: (1) Student-Teacher Cross-Attention Feature Fusion (STCA-FF), which adaptively integrates features from both models for a more interactive distillation process, and (2) Adaptive Spatial-Channel Masking (ASCM), which dynamically generates importance masks to enhance both spatial and channel-wise feature selection. Unlike conventional KD methods, ACAM-KD adapts to the student's evolving needs throughout the entire distillation process. Extensive experiments on multiple benchmarks validate its effectiveness. For instance, on COCO2017, ACAM-KD improves object detection performance by up to 1.4 mAP over the state-of-the-art when distilling a ResNet-50 student from a ResNet-101 teacher. For semantic segmentation on Cityscapes, it boosts mIoU by 3.09 over the baseline with DeepLabV3-MobileNetV2 as the student model.
In this work, we propose an extreme compression technique for Large Multimodal Models (LMMs). While previous studies have explored quantization as an efficient post-training compression method for Large Language Models (LLMs), low-bit compression for multimodal models remains under-explored. The redundant nature of inputs in multimodal models results in a highly sparse attention matrix. We theoretically and experimentally demonstrate that the attention matrix's sparsity bounds the compression error of the Query and Key weight matrices. Based on this, we introduce CASP, a model compression technique for LMMs. Our approach performs a data-aware low-rank decomposition on the Query and Key weight matrix, followed by quantization across all layers based on an optimal bit allocation process. CASP is compatible with any quantization technique and enhances state-of-the-art 2-bit quantization methods (AQLM and QuIP#) by an average of 21% on image- and video-language benchmarks.