Roaming in Wireless LAN (Wi-Fi) is a critical yet challenging task for maintaining seamless connectivity in dynamic mobile environments. Conventional threshold-based or heuristic schemes often fail, leading to either sticky or excessive handovers. We introduce the first cross-layer use of an on-device large language model (LLM): high-level reasoning in the application layer that issues real-time actions executed in the PHY/MAC stack. The LLM addresses two tasks: (i) context-aware AP selection, where structured prompts fuse environmental cues (e.g., location, time) to choose the best BSSID; and (ii) dynamic threshold adjustment, where the model adaptively decides when to roam. To satisfy the tight latency and resource budgets of edge hardware, we apply a suite of optimizations-chain-of-thought prompting, parameter-efficient fine-tuning, and quantization. Experiments on indoor and outdoor datasets show that our approach surpasses legacy heuristics and DRL baselines, achieving a strong balance between roaming stability and signal quality. These findings underscore the promise of application-layer LLM reasoning for lower-layer wireless control in future edge systems.
We introduce softpick, a rectified, not sum-to-one, drop-in replacement for softmax in transformer attention mechanisms that eliminates attention sink and massive activations. Our experiments with 340M parameter models demonstrate that softpick maintains performance parity with softmax on standard benchmarks while achieving 0% sink rate. The softpick transformer produces hidden states with significantly lower kurtosis (340 vs 33,510) and creates sparse attention maps (46.97% sparsity). Models using softpick consistently outperform softmax when quantized, with particularly pronounced advantages at lower bit precisions. Our analysis and discussion shows how softpick has the potential to open new possibilities for quantization, low-precision training, sparsity optimization, pruning, and interpretability. Our code is available at https://github.com/zaydzuhri/softpick-attention.
Autoregressive models have demonstrated remarkable success in sequential data generation, particularly in NLP, but their extension to continuous-domain image generation presents significant challenges. Recent work, the masked autoregressive model (MAR), bypasses quantization by modeling per-token distributions in continuous spaces using a diffusion head but suffers from slow inference due to the high computational cost of the iterative denoising process. To address this, we propose the Fast AutoRegressive model (FAR), a novel framework that replaces MAR's diffusion head with a lightweight shortcut head, enabling efficient few-step sampling while preserving autoregressive principles. Additionally, FAR seamlessly integrates with causal Transformers, extending them from discrete to continuous token generation without requiring architectural modifications. Experiments demonstrate that FAR achieves $2.3\times$ faster inference than MAR while maintaining competitive FID and IS scores. This work establishes the first efficient autoregressive paradigm for high-fidelity continuous-space image generation, bridging the critical gap between quality and scalability in visual autoregressive modeling.
Among the new techniques of Versatile Video Coding (VVC), the quadtree with nested multi-type tree (QT+MTT) block structure yields significant coding gains by providing more flexible block partitioning patterns. However, the recursive partition search in the VVC encoder increases the encoder complexity substantially. To address this issue, we propose a partition map-based algorithm to pursue fast block partitioning in inter coding. Based on our previous work on partition map-based methods for intra coding, we analyze the characteristics of VVC inter coding, and thus improve the partition map by incorporating an MTT mask for early termination. Next, we develop a neural network that uses both spatial and temporal features to predict the partition map. It consists of several special designs including stacked top-down and bottom-up processing, quantization parameter modulation layers, and partitioning-adaptive warping. Furthermore, we present a dual-threshold decision scheme to achieve a fine-grained trade-off between complexity reduction and rate-distortion (RD) performance loss. The experimental results demonstrate that the proposed method achieves an average 51.30% encoding time saving with a 2.12% Bjontegaard Delta Bit Rate (BDBR) under the random access configuration.
In modern large-scale systems with sensor networks and IoT devices it is essential to collaboratively solve complex problems while utilizing network resources efficiently. In our paper we present three distributed optimization algorithms that exhibit efficient communication among nodes. Our first algorithm presents a simple quantized averaged gradient procedure for distributed optimization, which is shown to converge to a neighborhood of the optimal solution. Our second algorithm incorporates a novel event-triggered refinement mechanism, which refines the utilized quantization level to enhance the precision of the estimated optimal solution. It enables nodes to terminate their operation according to predefined performance guarantees. Our third algorithm is tailored to operate in environments where each message consists of only a few bits. It incorporates a novel event-triggered mechanism for adjusting the quantizer basis and quantization level, allowing nodes to collaboratively decide operation termination based on predefined performance criteria. We analyze the three algorithms and establish their linear convergence. Finally, an application on distributed sensor fusion for target localization is used to demonstrate their favorable performance compared to existing algorithms in the literature.
Background: As empirical software engineering evolves, more studies adopt data strategies$-$approaches that investigate digital artifacts such as models, source code, or system logs rather than relying on human subjects. Synthesizing results from such studies introduces new methodological challenges. Aims: This study assesses the effects of model quantization on correctness and resource efficiency in deep learning (DL) systems. Additionally, it explores the methodological implications of aggregating evidence from empirical studies that adopt data strategies. Method: We conducted a research synthesis of six primary studies that empirically evaluate model quantization. We applied the Structured Synthesis Method (SSM) to aggregate the findings, which combines qualitative and quantitative evidence through diagrammatic modeling. A total of 19 evidence models were extracted and aggregated. Results: The aggregated evidence indicates that model quantization weakly negatively affects correctness metrics while consistently improving resource efficiency metrics, including storage size, inference latency, and GPU energy consumption$-$a manageable trade-off for many DL deployment contexts. Evidence across quantization techniques remains fragmented, underscoring the need for more focused empirical studies per technique. Conclusions: Model quantization offers substantial efficiency benefits with minor trade-offs in correctness, making it a suitable optimization strategy for resource-constrained environments. This study also demonstrates the feasibility of using SSM to synthesize findings from data strategy-based research.
Automatic speech quality assessment aims to quantify subjective human perception of speech through computational models to reduce the need for labor-consuming manual evaluations. While models based on deep learning have achieved progress in predicting mean opinion scores (MOS) to assess synthetic speech, the neglect of fundamental auditory perception mechanisms limits consistency with human judgments. To address this issue, we propose an auditory perception guided-MOS prediction model (APG-MOS) that synergistically integrates auditory modeling with semantic analysis to enhance consistency with human judgments. Specifically, we first design a perceptual module, grounded in biological auditory mechanisms, to simulate cochlear functions, which encodes acoustic signals into biologically aligned electrochemical representations. Secondly, we propose a residual vector quantization (RVQ)-based semantic distortion modeling method to quantify the degradation of speech quality at the semantic level. Finally, we design a residual cross-attention architecture, coupled with a progressive learning strategy, to enable multimodal fusion of encoded electrochemical signals and semantic representations. Experiments demonstrate that APG-MOS achieves superior performance on two primary benchmarks. Our code and checkpoint will be available on a public repository upon publication.
Federated learning enables decentralized model training while preserving data privacy, yet it faces challenges in balancing communication efficiency, model performance, and privacy protection. To address these trade-offs, we formulate FL as a federated multiobjective optimization problem and propose FedMOEAC, a clustering-based evolutionary algorithm that efficiently navigates the Pareto-optimal solution space. Our approach integrates quantization, weight sparsification, and differential privacy to reduce communication overhead while ensuring model robustness and privacy. The clustering mechanism en-hances population diversity, preventing premature convergence and improving optimization efficiency. Experimental results on MNIST and CIFAR-10 demonstrate that FedMOEAC achieves 98.2% accuracy, reduces communication overhead by 45%, and maintains a privacy budget below 1.0, outperforming NSGA-II in convergence speed by 33%. This work provides a scalable and efficient FL framework, ensuring an optimal balance between accuracy, communication efficiency, and privacy in resource-constrained environments.
Large Language Models (LLMs) have demonstrated remarkable capabilities in various natural language processing tasks. However, their size presents significant challenges for deployment and inference. This paper investigates the quantization of LLMs, focusing on the LLaMA architecture and its derivatives. We challenge existing assumptions about activation outliers in LLMs and propose a novel mixed-precision quantization approach tailored for LLaMA-like models. Our method leverages the observation that activation spikes in LLaMA architectures are predominantly concentrated in specific projection layers. By applying higher precision (FP16 or FP8) to these layers while quantizing the rest of the model to lower bit-widths, we achieve superior performance compared to existing quantization techniques. Experimental results on LLaMA2, LLaMA3, and Mistral models demonstrate significant improvements in perplexity and zero-shot accuracy, particularly for 8-bit per-tensor quantization. Our approach outperforms general-purpose methods designed to handle outliers across all architecture types, highlighting the benefits of architecture-specific quantization strategies. This research contributes to the ongoing efforts to make LLMs more efficient and deployable, potentially enabling their use in resource-constrained environments. Our findings emphasize the importance of considering model-specific characteristics in developing effective quantization pipelines for state-of-the-art language models by identifying and targeting a small number of projections that concentrate activation spikes.
Deep learning (DL) workloads mainly run on accelerators like GPUs. Recent DL quantization techniques demand a new matrix multiplication operator with mixed input data types, further complicating GPU optimization. Prior high-level compilers like Triton lack the expressiveness to implement key optimizations like fine-grained data pipelines and hardware-friendly memory layouts for these operators, while low-level programming models, such as Hidet, Graphene, and CUTLASS, require significant programming efforts. To balance expressiveness with engineering effort, we propose Hexcute, a tile-based programming language that exposes shared memory and register abstractions to enable fine-grained optimization for these operators. Additionally, Hexcute leverages task mapping to schedule the GPU program, and to reduce programming efforts, it automates layout and task mapping synthesis with a novel type-inference-based algorithm. Our evaluation shows that Hexcute generalizes to a wide range of DL operators, achieves 1.7-11.28$\times$ speedup over existing DL compilers for mixed-type operators, and brings up to 2.91$\times$ speedup in the end-to-end evaluation.
Quality of Experience (QoE) prediction plays a crucial role in optimizing resource management and enhancing user satisfaction across both telecommunication and OTT services. While recent advances predominantly rely on deep learning models, this study introduces a lightweight generative modeling framework that balances computational efficiency, interpretability, and predictive accuracy. By validating the use of Vector Quantization (VQ) as a preprocessing technique, continuous network features are effectively transformed into discrete categorical symbols, enabling integration with a Hidden Markov Model (HMM) for temporal sequence modeling. This VQ-HMM pipeline enhances the model's capacity to capture dynamic QoE patterns while supporting probabilistic inference on new and unseen data. Experimental results on publicly available time-series datasets incorporating both objective indicators and subjective QoE scores demonstrate the viability of this approach in real-time and resource-constrained environments, where inference latency is also critical. The framework offers a scalable alternative to complex deep learning methods, particularly in scenarios with limited computational resources or where latency constraints are critical.
Large language models (LLMs) have achieved remarkable success in contextual knowledge understanding. In this paper, we show that these concentrated massive values consistently emerge in specific regions of attention queries (Q) and keys (K) while not having such patterns in values (V) in various modern transformer-based LLMs (Q, K, and V mean the representations output by the query, key, and value layers respectively). Through extensive experiments, we further demonstrate that these massive values play a critical role in interpreting contextual knowledge (knowledge obtained from the current context window) rather than in retrieving parametric knowledge stored within the model's parameters. Our further investigation of quantization strategies reveals that ignoring these massive values leads to a pronounced drop in performance on tasks requiring rich contextual understanding, aligning with our analysis. Finally, we trace the emergence of concentrated massive values and find that such concentration is caused by Rotary Positional Encoding (RoPE), which has appeared since the first layers. These findings shed new light on how Q and K operate in LLMs and offer practical insights for model design and optimization. The Code is Available at https://github.com/MingyuJ666/Rope_with_LLM.
The Qwen series has emerged as a leading family of open-source Large Language Models (LLMs), demonstrating remarkable capabilities in natural language understanding tasks. With the recent release of Qwen3, which exhibits superior performance across diverse benchmarks, there is growing interest in deploying these models efficiently in resource-constrained environments. Low-bit quantization presents a promising solution, yet its impact on Qwen3's performance remains underexplored. This study conducts a systematic evaluation of Qwen3's robustness under various quantization settings, aiming to uncover both opportunities and challenges in compressing this state-of-the-art model. We rigorously assess 5 existing classic post-training quantization techniques applied to Qwen3, spanning bit-widths from 1 to 8 bits, and evaluate their effectiveness across multiple datasets. Our findings reveal that while Qwen3 maintains competitive performance at moderate bit-widths, it experiences notable degradation in linguistic tasks under ultra-low precision, underscoring the persistent hurdles in LLM compression. These results emphasize the need for further research to mitigate performance loss in extreme quantization scenarios. We anticipate that this empirical analysis will provide actionable insights for advancing quantization methods tailored to Qwen3 and future LLMs, ultimately enhancing their practicality without compromising accuracy. Our project is released on https://github.com/Efficient-ML/Qwen3-Quantization and https://huggingface.co/collections/Efficient-ML/qwen3-quantization-68164450decb1c868788cb2b.
In recent years, the compression of large language models (LLMs) has emerged as a key problem in facilitating LLM deployment on resource-limited devices, reducing compute costs, and mitigating the environmental footprint due to large-scale AI infrastructure. Here, we establish the foundations of LLM quantization from a rate-distortion theory perspective and propose a quantization technique based on simple rate-distortion optimization. Our technique scales to models containing hundreds of billions of weight parameters and offers users the flexibility to compress models, post-training, to a model size or accuracy specified by the user.
Large Language Models (LLMs) demonstrate exceptional performance across various tasks, but their large storage and computational requirements constrain their deployment on edge devices. To address this, we propose EntroLLM, a novel compression framework that integrates mixed quantization with entropy coding to reduce storage overhead while maintaining model accuracy. Our method applies a layer-wise mixed quantization scheme - choosing between symmetric and asymmetric quantization based on individual layer weight distributions - to optimize compressibility. We then employ Huffman encoding for lossless compression of the quantized weights, significantly reducing memory bandwidth requirements. Furthermore, we introduce parallel Huffman decoding, which enables efficient retrieval of encoded weights during inference, ensuring minimal latency impact. Our experiments on edge-compatible LLMs, including smolLM-1.7B-Instruct, phi3-mini-4k-Instruct, and mistral-7B-Instruct, demonstrate that EntroLLM achieves up to $30\%$ storage reduction compared to uint8 models and up to $65%$ storage reduction compared to uint4 models, while preserving perplexity and accuracy, on language benchmark tasks. We further show that our method enables $31.9\%$ - $146.6\%$ faster inference throughput on memory-bandwidth-limited edge devices, such as NVIDIA Jetson P3450, by reducing the required data movement. The proposed approach requires no additional re-training and is fully compatible with existing post-training quantization methods, making it a practical solution for edge LLMs.
Monocular depth estimation provides an additional depth dimension to RGB images, making it widely applicable in various fields such as virtual reality, autonomous driving and robotic navigation. However, existing depth estimation algorithms often struggle to effectively balance performance and computational efficiency, which poses challenges for deployment on resource-constrained devices. To address this, we propose LMDepth, a lightweight Mamba-based monocular depth estimation network, designed to reconstruct high-precision depth information while maintaining low computational overhead. Specifically, we propose a modified pyramid spatial pooling module that serves as a multi-scale feature aggregator and context extractor, ensuring global spatial information for accurate depth estimation. Moreover, we integrate multiple depth Mamba blocks into the decoder. Designed with linear computations, the Mamba Blocks enable LMDepth to efficiently decode depth information from global features, providing a lightweight alternative to Transformer-based architectures that depend on complex attention mechanisms. Extensive experiments on the NYUDv2 and KITTI datasets demonstrate the effectiveness of our proposed LMDepth. Compared to previous lightweight depth estimation methods, LMDepth achieves higher performance with fewer parameters and lower computational complexity (measured by GFLOPs). We further deploy LMDepth on an embedded platform with INT8 quantization, validating its practicality for real-world edge applications.
This paper presents a computationally efficient method for vehicle speed estimation from traffic camera footage. Building upon previous work that utilizes 3D bounding boxes derived from 2D detections and vanishing point geometry, we introduce several improvements to enhance real-time performance. We evaluate our method in several variants on the BrnoCompSpeed dataset in terms of vehicle detection and speed estimation accuracy. Our extensive evaluation across various hardware platforms, including edge devices, demonstrates significant gains in frames per second (FPS) compared to the prior state-of-the-art, while maintaining comparable or improved speed estimation accuracy. We analyze the trade-off between accuracy and computational cost, showing that smaller models utilizing post-training quantization offer the best balance for real-world deployment. Our best performing model beats previous state-of-the-art in terms of median vehicle speed estimation error (0.58 km/h vs. 0.60 km/h), detection precision (91.02% vs 87.08%) and recall (91.14% vs. 83.32%) while also being 5.5 times faster.
Diffusion models have recently emerged as the dominant approach in visual generation tasks. However, the lengthy denoising chains and the computationally intensive noise estimation networks hinder their applicability in low-latency and resource-limited environments. Previous research has endeavored to address these limitations in a decoupled manner, utilizing either advanced samplers or efficient model quantization techniques. In this study, we uncover that quantization-induced noise disrupts directional estimation at each sampling step, further distorting the precise directional estimations of higher-order samplers when solving the sampling equations through discretized numerical methods, thereby altering the optimal sampling trajectory. To attain dual acceleration with high fidelity, we propose a sampling-aware quantization strategy, wherein a Mixed-Order Trajectory Alignment technique is devised to impose a more stringent constraint on the error bounds at each sampling step, facilitating a more linear probability flow. Extensive experiments on sparse-step fast sampling across multiple datasets demonstrate that our approach preserves the rapid convergence characteristics of high-speed samplers while maintaining superior generation quality. Code will be made publicly available soon.
Cancer survival prediction using multi-modal medical imaging presents a critical challenge in oncology, mainly due to the vulnerability of deep learning models to noise and protocol variations across imaging centers. Current approaches struggle to extract consistent features from heterogeneous CT and PET images, limiting their clinical applicability. We address these challenges by introducing RobSurv, a robust deep-learning framework that leverages vector quantization for resilient multi-modal feature learning. The key innovation of our approach lies in its dual-path architecture: one path maps continuous imaging features to learned discrete codebooks for noise-resistant representation, while the parallel path preserves fine-grained details through continuous feature processing. This dual representation is integrated through a novel patch-wise fusion mechanism that maintains local spatial relationships while capturing global context via Transformer-based processing. In extensive evaluations across three diverse datasets (HECKTOR, H\&N1, and NSCLC Radiogenomics), RobSurv demonstrates superior performance, achieving concordance index of 0.771, 0.742, and 0.734 respectively - significantly outperforming existing methods. Most notably, our model maintains robust performance even under severe noise conditions, with performance degradation of only 3.8-4.5\% compared to 8-12\% in baseline methods. These results, combined with strong generalization across different cancer types and imaging protocols, establish RobSurv as a promising solution for reliable clinical prognosis that can enhance treatment planning and patient care.
The mixture of experts (MoE) model is a sparse variant of large language models (LLMs), designed to hold a better balance between intelligent capability and computational overhead. Despite its benefits, MoE is still too expensive to deploy on resource-constrained edge devices, especially with the demands of on-device inference services. Recent research efforts often apply model compression techniques, such as quantization, pruning and merging, to restrict MoE complexity. Unfortunately, due to their predefined static model optimization strategies, they cannot always achieve the desired quality-overhead trade-off when handling multiple requests, finally degrading the on-device quality of service. These limitations motivate us to propose the D$^2$MoE, an algorithm-system co-design framework that matches diverse task requirements by dynamically allocating the most proper bit-width to each expert. Specifically, inspired by the nested structure of matryoshka dolls, we propose the matryoshka weight quantization (MWQ) to progressively compress expert weights in a bit-nested manner and reduce the required runtime memory. On top of it, we further optimize the I/O-computation pipeline and design a heuristic scheduling algorithm following our hottest-expert-bit-first (HEBF) principle, which maximizes the expert parallelism between I/O and computation queue under constrained memory budgets, thus significantly reducing the idle temporal bubbles waiting for the experts to load. Evaluations on real edge devices show that D$^2$MoE improves the overall inference throughput by up to 1.39$\times$ and reduces the peak memory footprint by up to 53% over the latest on-device inference frameworks, while still preserving comparable serving accuracy as its INT8 counterparts.