We present MaskMark, a simple, efficient, and flexible framework for image watermarking. MaskMark has two variants: (1) MaskMark-D, which supports global watermark embedding, watermark localization, and local watermark extraction for applications such as tamper detection; (2) MaskMark-ED, which focuses on local watermark embedding and extraction, offering enhanced robustness in small regions to support fine-grined image protection. MaskMark-D builds on the classical encoder-distortion layer-decoder training paradigm. In MaskMark-D, we introduce a simple masking mechanism during the decoding stage that enables both global and local watermark extraction. During training, the decoder is guided by various types of masks applied to watermarked images before extraction, helping it learn to localize watermarks and extract them from the corresponding local areas. MaskMark-ED extends this design by incorporating the mask into the encoding stage as well, guiding the encoder to embed the watermark in designated local regions, which improves robustness under regional attacks. Extensive experiments show that MaskMark achieves state-of-the-art performance in global and local watermark extraction, watermark localization, and multi-watermark embedding. It outperforms all existing baselines, including the recent leading model WAM for local watermarking, while preserving high visual quality of the watermarked images. In addition, MaskMark is highly efficient and adaptable. It requires only 20 hours of training on a single A6000 GPU, achieving 15x computational efficiency compared to WAM. By simply adjusting the distortion layer, MaskMark can be quickly fine-tuned to meet varying robustness requirements.
Long-context language models exhibit impressive performance but remain challenging to deploy due to high GPU memory demands during inference. We propose Memory-efficient Offloaded Mini-sequence Inference (MOM), a method that partitions critical layers into smaller "mini-sequences" and integrates seamlessly with KV cache offloading. Experiments on various Llama, Qwen, and Mistral models demonstrate that MOM reduces peak memory usage by over 50\% on average. On Meta-Llama-3.2-8B, MOM extends the maximum context length from 155k to 455k tokens on a single A100 80GB GPU, while keeping outputs identical and not compromising accuracy. MOM also maintains highly competitive throughput due to minimal computational overhead and efficient last-layer processing. Compared to traditional chunked prefill methods, MOM achieves a 35\% greater context length extension. More importantly, our method drastically reduces prefill memory consumption, eliminating it as the longstanding dominant memory bottleneck during inference. This breakthrough fundamentally changes research priorities, redirecting future efforts from prefill-stage optimizations to improving decode-stage residual KV cache efficiency.
Clinical natural language processing (NLP) is increasingly in demand in both clinical research and operational practice. However, most of the state-of-the-art solutions are transformers-based and require high computational resources, limiting their accessibility. We propose a hybrid NLP framework that integrates rule-based filtering, a Support Vector Machine (SVM) classifier, and a BERT-based model to improve efficiency while maintaining accuracy. We applied this framework in a dementia identification case study involving 4.9 million veterans with incident hypertension, analyzing 2.1 billion clinical notes. At the patient level, our method achieved a precision of 0.90, a recall of 0.84, and an F1-score of 0.87. Additionally, this NLP approach identified over three times as many dementia cases as structured data methods. All processing was completed in approximately two weeks using a single machine with dual A40 GPUs. This study demonstrates the feasibility of hybrid NLP solutions for large-scale clinical text analysis, making state-of-the-art methods more accessible to healthcare organizations with limited computational resources.
This paper introduces ConvShareViT, a novel deep learning architecture that adapts Vision Transformers (ViTs) to the 4f free-space optical system. ConvShareViT replaces linear layers in multi-head self-attention (MHSA) and Multilayer Perceptrons (MLPs) with a depthwise convolutional layer with shared weights across input channels. Through the development of ConvShareViT, the behaviour of convolutions within MHSA and their effectiveness in learning the attention mechanism were analysed systematically. Experimental results demonstrate that certain configurations, particularly those using valid-padded shared convolutions, can successfully learn attention, achieving comparable attention scores to those obtained with standard ViTs. However, other configurations, such as those using same-padded convolutions, show limitations in attention learning and operate like regular CNNs rather than transformer models. ConvShareViT architectures are specifically optimised for the 4f optical system, which takes advantage of the parallelism and high-resolution capabilities of optical systems. Results demonstrate that ConvShareViT can theoretically achieve up to 3.04 times faster inference than GPU-based systems. This potential acceleration makes ConvShareViT an attractive candidate for future optical deep learning applications and proves that our ViT (ConvShareViT) can be employed using only the convolution operation, via the necessary optimisation of the ViT to balance performance and complexity.
Serving Large Language Models (LLMs) is critical for AI-powered applications but demands substantial computational resources, particularly in memory bandwidth and computational throughput. Low-precision computation has emerged as a key technique to improve efficiency while reducing resource consumption. Existing approaches for generating low-precision kernels are limited to weight bit widths that are powers of two and suffer from suboptimal performance due to high-level GPU programming abstractions. These abstractions restrict critical optimizations, such as fine-grained register management and optimized memory access patterns, which are essential for efficient low-precision computations. In this paper, we introduce a virtual machine (VM) designed for General-Purpose GPU (GPGPU) computing, enabling support for low-precision data types with arbitrary bit widths while maintaining GPU programmability. The proposed VM features a thread-block-level programming model, a hierarchical memory space, a novel algebraic layout system, and extensive support for diverse low-precision data types. VM programs are compiled into highly efficient GPU programs with automatic vectorization and instruction selection. Extensive experiments demonstrate that our VM efficiently supports a full spectrum of low-precision data types, and outperforms state-of-the-art low-precision kernels on their supported types. Compared to existing compilers like Triton and Ladder, as well as hand-optimized kernels such as QuantLLM and Marlin, our VM achieves performance improvements of 1.75x, 2.61x, 1.29x and 1.03x, respectively.
Drag-driven editing has become popular among designers for its ability to modify complex geometric structures through simple and intuitive manipulation, allowing users to adjust and reshape content with minimal technical skill. This drag operation has been incorporated into numerous methods to facilitate the editing of 2D images and 3D meshes in design. However, few studies have explored drag-driven editing for the widely-used 3D Gaussian Splatting (3DGS) representation, as deforming 3DGS while preserving shape coherence and visual continuity remains challenging. In this paper, we introduce ARAP-GS, a drag-driven 3DGS editing framework based on As-Rigid-As-Possible (ARAP) deformation. Unlike previous 3DGS editing methods, we are the first to apply ARAP deformation directly to 3D Gaussians, enabling flexible, drag-driven geometric transformations. To preserve scene appearance after deformation, we incorporate an advanced diffusion prior for image super-resolution within our iterative optimization process. This approach enhances visual quality while maintaining multi-view consistency in the edited results. Experiments show that ARAP-GS outperforms current methods across diverse 3D scenes, demonstrating its effectiveness and superiority for drag-driven 3DGS editing. Additionally, our method is highly efficient, requiring only 10 to 20 minutes to edit a scene on a single RTX 3090 GPU.
LLM inference is essential for applications like text summarization, translation, and data analysis, but the high cost of GPU instances from Cloud Service Providers (CSPs) like AWS is a major burden. This paper proposes InferSave, a cost-efficient VM selection framework for cloud based LLM inference. InferSave optimizes KV cache offloading based on Service Level Objectives (SLOs) and workload charac teristics, estimating GPU memory needs, and recommending cost-effective VM instances. Additionally, the Compute Time Calibration Function (CTCF) improves instance selection accuracy by adjusting for discrepancies between theoretical and actual GPU performance. Experiments on AWS GPU instances show that selecting lower-cost instances without KV cache offloading improves cost efficiency by up to 73.7% for online workloads, while KV cache offloading saves up to 20.19% for offline workloads.
Large Language Models (LLMs) have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM size by 30% while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) decomposition of memory-intensive lookup tables (LUTs) into compact LUTs that fit in GPU SRAM, (ii) a two-phase kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on recent models, including Llama-3.1, Qwen-2.5, and Gemma-3, validates our hypothesis that DFloat11 achieves around 30% model size reduction while preserving bit-for-bit exact outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves 1.9-38.8x higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables 5.3-13.17x longer context lengths than uncompressed models. Notably, our method enables lossless inference of Llama-3.1-405B, an 810GB model, on a single node equipped with 8x80GB GPUs. Our code and models are available at https://github.com/LeanModels/DFloat11.
Tactile sensing is crucial for achieving human-level robotic capabilities in manipulation tasks. VBTSs have emerged as a promising solution, offering high spatial resolution and cost-effectiveness by sensing contact through camera-captured deformation patterns of elastic gel pads. However, these sensors' complex physical characteristics and visual signal processing requirements present unique challenges for robotic applications. The lack of efficient and accurate simulation tools for VBTS has significantly limited the scale and scope of tactile robotics research. Here we present Taccel, a high-performance simulation platform that integrates IPC and ABD to model robots, tactile sensors, and objects with both accuracy and unprecedented speed, achieving an 18-fold acceleration over real-time across thousands of parallel environments. Unlike previous simulators that operate at sub-real-time speeds with limited parallelization, Taccel provides precise physics simulation and realistic tactile signals while supporting flexible robot-sensor configurations through user-friendly APIs. Through extensive validation in object recognition, robotic grasping, and articulated object manipulation, we demonstrate precise simulation and successful sim-to-real transfer. These capabilities position Taccel as a powerful tool for scaling up tactile robotics research and development. By enabling large-scale simulation and experimentation with tactile sensing, Taccel accelerates the development of more capable robotic systems, potentially transforming how robots interact with and understand their physical environment.
Large language models (LLMs) have demonstrated significant potential in code generation tasks. However, there remains a performance gap between open-source and closed-source models. To address this gap, existing approaches typically generate large amounts of synthetic data for fine-tuning, which often leads to inefficient training. In this work, we propose a data selection strategy in order to improve the effectiveness and efficiency of training for code-based LLMs. By prioritizing data complexity and ensuring that the sampled subset aligns with the distribution of the original dataset, our sampling strategy effectively selects high-quality data. Additionally, we optimize the tokenization process through a "dynamic pack" technique, which minimizes padding tokens and reduces computational resource consumption. Experimental results show that when training on 40% of the OSS-Instruct dataset, the DeepSeek-Coder-Base-6.7B model achieves an average performance of 66.9%, surpassing the 66.1% performance with the full dataset. Moreover, training time is reduced from 47 minutes to 34 minutes, and the peak GPU memory decreases from 61.47 GB to 42.72 GB during a single epoch. Similar improvements are observed with the CodeLlama-Python-7B model on the Evol-Instruct dataset. By optimizing both data selection and tokenization, our approach not only improves model performance but also improves training efficiency.
Large language model (LLM)-based inference workloads increasingly dominate data center costs and resource utilization. Therefore, understanding the inference workload characteristics on evolving CPU-GPU coupled architectures is crucial for optimization. This paper presents an in-depth analysis of LLM inference behavior on loosely-coupled (PCIe A100/H100) and closely-coupled (GH200) systems. We analyze performance dynamics using fine-grained operator-to-kernel trace analysis, facilitated by our novel profiler SKIP and metrics like Total Kernel Launch and Queuing Time (TKLQT). Results show that closely-coupled (CC) GH200 significantly outperforms loosely-coupled (LC) systems at large batch sizes, achieving 1.9x-2.7x faster prefill latency for Llama 3.2-1B. However, our analysis also reveals that GH200 remains CPU-bound up to 4x larger batch sizes than LC systems. In this extended CPU-bound region, we identify the performance characteristics of the Grace CPU as a key factor contributing to higher inference latency at low batch sizes on GH200. We demonstrate that TKLQT accurately identifies this CPU/GPU-bound transition point. Based on this analysis, we further show that kernel fusion offers significant potential to mitigate GH200's low-batch latency bottleneck by reducing kernel launch overhead. This detailed kernel-level characterization provides critical insights for optimizing diverse CPU-GPU coupling strategies. This work is an initial effort, and we plan to explore other major AI/DL workloads that demand different degrees of CPU-GPU heterogeneous architectures.
Global contextual information and local detail features are essential for haze removal tasks. Deep learning models perform well on small, low-resolution images, but they encounter difficulties with large, high-resolution ones due to GPU memory limitations. As a compromise, they often resort to image slicing or downsampling. The former diminishes global information, while the latter discards high-frequency details. To address these challenges, we propose DehazeXL, a haze removal method that effectively balances global context and local feature extraction, enabling end-to-end modeling of large images on mainstream GPU hardware. Additionally, to evaluate the efficiency of global context utilization in haze removal performance, we design a visual attribution method tailored to the characteristics of haze removal tasks. Finally, recognizing the lack of benchmark datasets for haze removal in large images, we have developed an ultra-high-resolution haze removal dataset (8KDehaze) to support model training and testing. It includes 10000 pairs of clear and hazy remote sensing images, each sized at 8192 $\times$ 8192 pixels. Extensive experiments demonstrate that DehazeXL can infer images up to 10240 $\times$ 10240 pixels with only 21 GB of memory, achieving state-of-the-art results among all evaluated methods. The source code and experimental dataset are available at https://github.com/CastleChen339/DehazeXL.
Pansharpening is a crucial remote sensing technique that fuses low-resolution multispectral (LRMS) images with high-resolution panchromatic (PAN) images to generate high-resolution multispectral (HRMS) imagery. Although deep learning techniques have significantly advanced pansharpening, many existing methods suffer from limited cross-sensor generalization and high computational overhead, restricting their real-time applications. To address these challenges, we propose an efficient framework that quickly adapts to a specific input instance, completing both training and inference in a short time. Our framework splits the input image into multiple patches, selects a subset for unsupervised CAT training, and then performs inference on all patches, stitching them into the final output. The CAT module, integrated between the feature extraction and channel transformation stages of a pre-trained network, tailors the fused features and fixes the parameters for efficient inference, generating improved results. Our approach offers two key advantages: (1) $\textit{Improved Generalization Ability}$: by mitigating cross-sensor degradation, our model--although pre-trained on a specific dataset--achieves superior performance on datasets captured by other sensors; (2) $\textit{Enhanced Computational Efficiency}$: the CAT-enhanced network can swiftly adapt to the test sample using the single LRMS-PAN pair input, without requiring extensive large-scale data retraining. Experiments on the real-world data from WorldView-3 and WorldView-2 datasets demonstrate that our method achieves state-of-the-art performance on cross-sensor real-world data, while achieving both training and inference of $512\times512$ image within $\textit{0.4 seconds}$ and $4000\times4000$ image within $\textit{3 seconds}$ at the fastest setting on a commonly used RTX 3090 GPU.
Vision foundation models (VFMs) are large pre-trained models that form the backbone of various vision tasks. Fine-tuning VFMs can further unlock their potential for downstream tasks or scenarios. However, VFMs often contain significant feature redundancy, which may limit their adaptability to new tasks. In this paper, we investigate the redundancies in the segment anything model (SAM) and then propose a parameter-free fine-tuning method to address this issue. Unlike traditional fine-tuning methods that adjust parameters, our method emphasizes selecting, reusing, and enhancing pre-trained features, offering a new perspective on model fine-tuning. Specifically, we introduce a channel selection algorithm based on the model's output difference to identify redundant and effective channels. By selectively replacing the redundant channels with more effective ones, we filter out less useful features and reuse the more relevant features to downstream tasks, thereby enhancing the task-specific feature representation. Experiments on both out-of-domain and in-domain datasets demonstrate the efficiency and effectiveness of our method. Notably, our approach can seamlessly integrate with existing fine-tuning strategies (e.g., LoRA, Adapter), further boosting the performance of already fine-tuned models. Moreover, since our channel selection involves only model inference, our method significantly reduces computational and GPU memory overhead.
Diffusion models have recently gained state of the art performance on many image generation tasks. However, most models require significant computational resources to achieve this. This becomes apparent in the application of medical image synthesis due to the 3D nature of medical datasets like CT-scans, MRIs, electron microscope, etc. In this paper we propose a novel architecture for a single GPU memory-efficient training for diffusion models for high dimensional medical datasets. The proposed model is built by using an invertible UNet architecture with invertible attention modules. This leads to the following two contributions: 1. denoising diffusion models and thus enabling memory usage to be independent of the dimensionality of the dataset, and 2. reducing the energy usage during training. While this new model can be applied to a multitude of image generation tasks, we showcase its memory-efficiency on the 3D BraTS2020 dataset leading to up to 15\% decrease in peak memory consumption during training with comparable results to SOTA while maintaining the image quality.
Increasing attention has been placed on improving the reasoning capacities of multi-modal large language models (MLLMs). As the cornerstone for AI agents that function in the physical realm, video-based visual-spatial intelligence (VSI) emerges as one of the most pivotal reasoning capabilities of MLLMs. This work conducts a first, in-depth study on improving the visual-spatial reasoning of MLLMs via R1-Zero-like training. Technically, we first identify that the visual-spatial reasoning capacities of small- to medium-sized Qwen2-VL models cannot be activated via Chain of Thought (CoT) prompts. We then incorporate GRPO training for improved visual-spatial reasoning, using the carefully curated VSI-100k dataset, following DeepSeek-R1-Zero. During the investigation, we identify the necessity to keep the KL penalty (even with a small value) in GRPO. With just 120 GPU hours, our vsGRPO-2B model, fine-tuned from Qwen2-VL-2B, can outperform the base model by 12.1% and surpass GPT-4o. Moreover, our vsGRPO-7B model, fine-tuned from Qwen2-VL-7B, achieves performance comparable to that of the best open-source model LLaVA-NeXT-Video-72B. Additionally, we compare vsGRPO to supervised fine-tuning and direct preference optimization baselines and observe strong performance superiority. The code and dataset will be available soon.
Adoption of AI driven techniques in public services remains low due to challenges related to accuracy and speed of information at population scale. Computer vision techniques for traffic monitoring have not gained much popularity despite their relative strength in areas such as autonomous driving. Despite large number of academic methods for Automatic License Plate Recognition (ALPR) systems, very few provide an end to end solution for patrolling in the city. This paper presents a novel prototype for a low power GPU based patrolling system to be deployed in an urban environment on surveillance vehicles for automated vehicle detection, recognition and tracking. In this work, we propose a complete ALPR system for Singapore license plates having both single and double line creating our own YOLO based network. We focus on unconstrained capture scenarios as would be the case in real world application, where the license plate (LP) might be considerably distorted due to oblique views. In this work, we first detect the license plate from the full image using RFB-Net and rectify multiple distorted license plates in a single image. After that, the detected license plate image is fed to our network for character recognition. We evaluate the performance of our proposed system on a newly built dataset covering more than 16,000 images. The system was able to correctly detect license plates with 86\% precision and recognize characters of a license plate in 67\% of the test set, and 89\% accuracy with one incorrect character (partial match). We also test latency of our system and achieve 64FPS on Tesla P4 GPU
As model sizes grow, finding efficient and cost-effective hyperparameter optimization (HPO) methods becomes increasingly crucial for deep learning pipelines. While multi-fidelity HPO (MF-HPO) trades off computational resources required for DL training with lower fidelity estimations, existing fidelity sources often fail under lower compute and memory constraints. We propose a novel fidelity source: the number of layers that are trained or frozen during training. For deep networks, this approach offers significant compute and memory savings while preserving rank correlations between hyperparameters at low fidelities compared to full model training. We demonstrate this in our empirical evaluation across ResNets and Transformers and additionally analyze the utility of frozen layers as a fidelity in using GPU resources as a fidelity in HPO, and for a combined MF-HPO with other fidelity sources. This contribution opens new applications for MF-HPO with hardware resources as a fidelity and creates opportunities for improved algorithms navigating joint fidelity spaces.
Deep neural networks (DNNs) enhance the accuracy and efficiency of reconstructing key parameters from time-resolved photon arrival signals recorded by single-photon detectors. However, the performance of conventional backpropagation-based DNNs is highly dependent on various parameters of the optical setup and biological samples under examination, necessitating frequent network retraining, either through transfer learning or from scratch. Newly collected data must also be stored and transferred to a high-performance GPU server for retraining, introducing latency and storage overhead. To address these challenges, we propose an online training algorithm based on a One-Sided Jacobi rotation-based Online Sequential Extreme Learning Machine (OSOS-ELM). We fully exploit parallelism in executing OSOS-ELM on a heterogeneous FPGA with integrated ARM cores. Extensive evaluations of OSOS-ELM and OSELM demonstrate that both achieve comparable accuracy across different network dimensions (i.e., input, hidden, and output layers), while OSOS-ELM proves to be more hardware-efficient. By leveraging the parallelism of OSOS-ELM, we implement a holistic computing prototype on a Xilinx ZCU104 FPGA, which integrates a multi-core CPU and programmable logic fabric. We validate our approach through three case studies involving single-photon signal analysis: sensing through fog using commercial single-photon LiDAR, fluorescence lifetime estimation in FLIM, and blood flow index reconstruction in DCS, all utilizing one-dimensional data encoded from photonic signals. From a hardware perspective, we optimize the OSOS-ELM workload by employing multi-tasked processing on ARM CPU cores and pipelined execution on the FPGA's logic fabric. We also implement our OSOS-ELM on the NVIDIA Jetson Xavier NX GPU to comprehensively investigate its computing performance on another type of heterogeneous computing platform.
We develop a novel approach to constrain the Hubble parameter $H_0$ and the primordial power spectrum amplitude $A_\mathrm{s}$ using supernovae type Ia (SNIa) data. By considering SNIa as tracers of the peculiar velocity field, we can model their distance and their covariance as a function of cosmological parameters without the need of calibrators like Cepheids; this yields a new independent probe of the large-scale structure based on SNIa data without distance anchors. Crucially, we implement a differentiable pipeline in JAX, including efficient emulators and affine sampling, reducing inference time from years to hours on a single GPU. We first validate our method on mock datasets, demonstrating that we can constrain $H_0$ and $\log 10^{10}A_\mathrm{s}$ within $\sim10\%$ using $\sim10^3$ SNIa. We then test our pipeline with SNIa from an $N$-body simulation, obtaining $7\%$-level unbiased constraints on $H_0$ with a moderate noise level. We finally apply our method to Pantheon+ data, constraining $H_0$ at the $10\%$ level without Cepheids when fixing $A_\mathrm{s}$ to its $\it{Planck}$ value. On the other hand, we obtain $15\%$-level constraints on $\log 10^{10}A_\mathrm{s}$ in agreement with $\it{Planck}$ when including Cepheids in the analysis. In light of upcoming observations of low redshift SNIa from the Zwicky Transient Facility and the Vera Rubin Legacy Survey of Space and Time, surveys for which our method will develop its full potential, we make our code publicly available.