Developing a contemporary optimal transport (OT) solver requires navigating trade-offs among several critical requirements: GPU parallelization, scalability to high-dimensional problems, theoretical convergence guarantees, empirical performance in terms of precision versus runtime, and numerical stability in practice. With these challenges in mind, we introduce a specialized truncated Newton algorithm for entropic-regularized OT. In addition to proving that locally quadratic convergence is possible without assuming a Lipschitz Hessian, we provide strategies to maximally exploit the high rate of local convergence in practice. Our GPU-parallel algorithm exhibits exceptionally favorable runtime performance, achieving high precision orders of magnitude faster than many existing alternatives. This is evidenced by wall-clock time experiments on 24 problem sets (12 datasets $\times$ 2 cost functions). The scalability of the algorithm is showcased on an extremely large OT problem with $n \approx 10^6$, solved approximately under weak entopric regularization.
The reproduction of state-of-the-art multimodal LLM pre-training faces barriers at every stage of the pipeline, including high-quality data filtering, multimodal data mixture strategies, sequence packing techniques, and training frameworks. We introduce Open-Qwen2VL, a fully open-source 2B-parameter Multimodal Large Language Model pre-trained efficiently on 29M image-text pairs using only 220 A100-40G GPU hours. Our approach employs low-to-high dynamic image resolution and multimodal sequence packing to significantly enhance pre-training efficiency. The training dataset was carefully curated using both MLLM-based filtering techniques (e.g., MLM-Filter) and conventional CLIP-based filtering methods, substantially improving data quality and training efficiency. The Open-Qwen2VL pre-training is conducted on academic level 8xA100-40G GPUs at UCSB on 5B packed multimodal tokens, which is 0.36% of 1.4T multimodal pre-training tokens of Qwen2-VL. The final instruction-tuned Open-Qwen2VL outperforms partially-open state-of-the-art MLLM Qwen2-VL-2B on various multimodal benchmarks of MMBench, SEEDBench, MMstar, and MathVista, indicating the remarkable training efficiency of Open-Qwen2VL. We open-source all aspects of our work, including compute-efficient and data-efficient training details, data filtering methods, sequence packing scripts, pre-training data in WebDataset format, FSDP-based training codebase, and both base and instruction-tuned model checkpoints. We redefine "fully open" for multimodal LLMs as the complete release of: 1) the training codebase, 2) detailed data filtering techniques, and 3) all pre-training and supervised fine-tuning data used to develop the model.
3D Gaussian splatting enables high-quality novel view synthesis (NVS) at real-time frame rates. However, its quality drops sharply as we depart from the training views. Thus, dense captures are needed to match the high-quality expectations of some applications, e.g. Virtual Reality (VR). However, such dense captures are very laborious and expensive to obtain. Existing works have explored using 2D generative models to alleviate this requirement by distillation or generating additional training views. These methods are often conditioned only on a handful of reference input views and thus do not fully exploit the available 3D information, leading to inconsistent generation results and reconstruction artifacts. To tackle this problem, we propose a multi-view, flow matching model that learns a flow to connect novel view renderings from possibly sparse reconstructions to renderings that we expect from dense reconstructions. This enables augmenting scene captures with novel, generated views to improve reconstruction quality. Our model is trained on a novel dataset of 3.6M image pairs and can process up to 45 views at 540x960 resolution (91K tokens) on one H100 GPU in a single forward pass. Our pipeline consistently improves NVS in sparse- and dense-view scenarios, leading to higher-quality reconstructions than prior works across multiple, widely-used NVS benchmarks.
Recent advancements in large language models (LLMs) have prompted interest in deploying these models on mobile devices to enable new applications without relying on cloud connectivity. However, the efficiency constraints of deploying LLMs on resource-limited devices present significant challenges. In this paper, we conduct a comprehensive measurement study to evaluate the efficiency tradeoffs between mobile-based, edge-based, and cloud-based deployments for LLM applications. We implement AutoLife-Lite, a simplified LLM-based application that analyzes smartphone sensor data to infer user location and activity contexts. Our experiments reveal that: (1) Only small-size LLMs (<4B parameters) can run successfully on powerful mobile devices, though they exhibit quality limitations compared to larger models; (2) Model compression is effective in lower the hardware requirement, but may lead to significant performance degradation; (3) The latency to run LLMs on mobile devices with meaningful output is significant (>30 seconds), while cloud services demonstrate better time efficiency (<10 seconds); (4) Edge deployments offer intermediate tradeoffs between latency and model capabilities, with different results on CPU-based and GPU-based settings. These findings provide valuable insights for system designers on the current limitations and future directions for on-device LLM applications.
The Internet of Vehicles (IoV) may face challenging cybersecurity attacks that may require sophisticated intrusion detection systems, necessitating a rapid development and response system. This research investigates the performance advantages of GPU-accelerated libraries (cuML) compared to traditional CPU-based implementations (scikit-learn), focusing on the speed and efficiency required for machine learning models used in IoV threat detection environments. The comprehensive evaluations conducted employ four machine learning approaches (Random Forest, KNN, Logistic Regression, XGBoost) across three distinct IoV security datasets (OTIDS, GIDS, CICIoV2024). Our findings demonstrate that GPU-accelerated implementations dramatically improved computational efficiency, with training times reduced by a factor of up to 159 and prediction speeds accelerated by up to 95 times compared to traditional CPU processing, all while preserving detection accuracy. This remarkable performance breakthrough empowers researchers and security specialists to harness GPU acceleration for creating faster, more effective threat detection systems that meet the urgent real-time security demands of today's connected vehicle networks.
Real-time optimal control remains a fundamental challenge in robotics, especially for nonlinear systems with stringent performance requirements. As one of the representative trajectory optimization algorithms, the iterative Linear Quadratic Regulator (iLQR) faces limitations due to their inherently sequential computational nature, which restricts the efficiency and applicability of real-time control for robotic systems. While existing parallel implementations aim to overcome the above limitations, they typically demand additional computational iterations and high-performance hardware, leading to only modest practical improvements. In this paper, we introduce Quattro, a transformer-accelerated iLQR framework employing an algorithm-hardware co-design strategy to predict intermediate feedback and feedforward matrices. It facilitates effective parallel computations on resource-constrained devices without sacrificing accuracy. Experiments on cart-pole and quadrotor systems show an algorithm-level acceleration of up to 5.3$\times$ and 27$\times$ per iteration, respectively. When integrated into a Model Predictive Control (MPC) framework, Quattro achieves overall speedups of 2.8$\times$ for the cart-pole and 17.8$\times$ for the quadrotor compared to the one that applies traditional iLQR. Transformer inference is deployed on FPGA to maximize performance, achieving further up to 20.8$\times$ speedup over prevalent embedded CPUs with over 11$\times$ power reduction than GPU and low hardware resource overhead.
Deep Learning Recommendation Models (DLRMs) play a crucial role in delivering personalized content across web applications such as social networking and video streaming. However, with improvements in performance, the parameter size of DLRMs has grown to terabyte (TB) scales, accompanied by memory bandwidth demands exceeding TB/s levels. Furthermore, the workload intensity within the model varies based on the target mechanism, making it difficult to build an optimized recommendation system. In this paper, we propose SCRec, a scalable computational storage recommendation system that can handle TB-scale industrial DLRMs while guaranteeing high bandwidth requirements. SCRec utilizes a software framework that features a mixed-integer programming (MIP)-based cost model, efficiently fetching data based on data access patterns and adaptively configuring memory-centric and compute-centric cores. Additionally, SCRec integrates hardware acceleration cores to enhance DLRM computations, particularly allowing for the high-performance reconstruction of approximated embedding vectors from extremely compressed tensor-train (TT) format. By combining its software framework and hardware accelerators, while eliminating data communication overhead by being implemented on a single server, SCRec achieves substantial improvements in DLRM inference performance. It delivers up to 55.77$\times$ speedup compared to a CPU-DRAM system with no loss in accuracy and up to 13.35$\times$ energy efficiency gains over a multi-GPU system.
As the demands for railway transportation safety increase, traditional methods of rail track inspection no longer meet the needs of modern railway systems. To address the issues of automation and efficiency in rail fault detection, this study introduces a railway inspection system based on Field Programmable Gate Array (FPGA). This edge AI system collects track images via cameras and uses Convolutional Neural Networks (CNN) to perform real-time detection of track defects and automatically reports fault information. The innovation of this system lies in its high level of automation and detection efficiency. The neural network approach employed by this system achieves a detection accuracy of 88.9%, significantly enhancing the reliability and efficiency of detection. Experimental results demonstrate that this FPGA-based system is 1.39* and 4.67* better in energy efficiency than peer implementation on the GPU and CPU platform, respectively.
Artificial intelligence has progressed through the development of Vision-Language Models (VLMs), which integrate text and visual inputs to achieve comprehensive understanding and interaction in various contexts. Enhancing the performance of these models such as the transformer based Florence 2 on specialized tasks like object detection in complex and unstructured environments requires fine-tuning. The goal of this paper is to improve the efficiency of the Florence 2 model in challenging environments by finetuning it. We accomplished this by experimenting with different configurations, using various GPU types (T4, L4, A100) and optimizers such as AdamW and SGD. We also employed a range of learning rates and LoRA (Low Rank Adaptation) settings. Analyzing the performance metrics, such as Mean Average Precision (mAP) scores,reveals that the finetuned Florence 2 models performed comparably to YOLO models, including YOLOv8, YOLOv9, and YOLOv10. This demonstrates how transformer based VLMs can be adapted for detailed object detection tasks. The paper emphasizes the capability of optimized transformer based VLMs to address specific challenges in object detection within unstructured environments, opening up promising avenues for practical applications in demanding and complex settings.
Visual Autoregressive (VAR) modeling has gained popularity for its shift towards next-scale prediction. However, existing VAR paradigms process the entire token map at each scale step, leading to the complexity and runtime scaling dramatically with image resolution. To address this challenge, we propose FastVAR, a post-training acceleration method for efficient resolution scaling with VARs. Our key finding is that the majority of latency arises from the large-scale step where most tokens have already converged. Leveraging this observation, we develop the cached token pruning strategy that only forwards pivotal tokens for scale-specific modeling while using cached tokens from previous scale steps to restore the pruned slots. This significantly reduces the number of forwarded tokens and improves the efficiency at larger resolutions. Experiments show the proposed FastVAR can further speedup FlashAttention-accelerated VAR by 2.7$\times$ with negligible performance drop of <1%. We further extend FastVAR to zero-shot generation of higher resolution images. In particular, FastVAR can generate one 2K image with 15GB memory footprints in 1.5s on a single NVIDIA 3090 GPU. Code is available at https://github.com/csguoh/FastVAR.
Recently, large language models (LLMs) have been able to handle longer and longer contexts. However, a context that is too long may cause intolerant inference latency and GPU memory usage. Existing methods propose mixed-precision quantization to the key-value (KV) cache in LLMs based on token granularity, which is time-consuming in the search process and hardware inefficient during computation. This paper introduces a novel approach called Cocktail, which employs chunk-adaptive mixed-precision quantization to optimize the KV cache. Cocktail consists of two modules: chunk-level quantization search and chunk-level KV cache computation. Chunk-level quantization search determines the optimal bitwidth configuration of the KV cache chunks quickly based on the similarity scores between the corresponding context chunks and the query, maintaining the model accuracy. Furthermore, chunk-level KV cache computation reorders the KV cache chunks before quantization, avoiding the hardware inefficiency caused by mixed-precision quantization in inference computation. Extensive experiments demonstrate that Cocktail outperforms state-of-the-art KV cache quantization methods on various models and datasets.
Compute nodes on modern heterogeneous supercomputing systems comprise CPUs, GPUs, and high-speed network interconnects (NICs). Parallelization is identified as a technique for effectively utilizing these systems to execute scalable simulation and deep learning workloads. The resulting inter-process communication from the distributed execution of these parallel workloads is one of the key factors contributing to its performance bottleneck. Most programming models and runtime systems enabling the communication requirements on these systems support GPU-aware communication schemes that move the GPU-attached communication buffers in the application directly from the GPU to the NIC without staging through the host memory. A CPU thread is required to orchestrate the communication operations even with support for such GPU-awareness. This survey discusses various available GPU-centric communication schemes that move the control path of the communication operations from the CPU to the GPU. This work presents the need for the new communication schemes, various GPU and NIC capabilities required to implement the schemes, and the potential use-cases addressed. Based on these discussions, challenges involved in supporting the exhibited GPU-centric communication schemes are discussed.
On-device deep learning (DL) has rapidly gained adoption in mobile apps, offering the benefits of offline model inference and user privacy preservation over cloud-based approaches. However, it inevitably stores models on user devices, introducing new vulnerabilities, particularly model-stealing attacks and intellectual property infringement. While system-level protections like Trusted Execution Environments (TEEs) provide a robust solution, practical challenges remain in achieving scalable on-device DL model protection, including complexities in supporting third-party models and limited adoption in current mobile solutions. Advancements in TEE-enabled hardware, such as NVIDIA's GPU-based TEEs, may address these obstacles in the future. Currently, watermarking serves as a common defense against model theft but also faces challenges here as many mobile app developers lack corresponding machine learning expertise and the inherent read-only and inference-only nature of on-device DL models prevents third parties like app stores from implementing existing watermarking techniques in post-deployment models. To protect the intellectual property of on-device DL models, in this paper, we propose THEMIS, an automatic tool that lifts the read-only restriction of on-device DL models by reconstructing their writable counterparts and leverages the untrainable nature of on-device DL models to solve watermark parameters and protect the model owner's intellectual property. Extensive experimental results across various datasets and model structures show the superiority of THEMIS in terms of different metrics. Further, an empirical investigation of 403 real-world DL mobile apps from Google Play is performed with a success rate of 81.14%, showing the practicality of THEMIS.
Recently, deep learning-based pan-sharpening algorithms have achieved notable advancements over traditional methods. However, many deep learning-based approaches incur substantial computational overhead during inference, especially with high-resolution images. This excessive computational demand limits the applicability of these methods in real-world scenarios, particularly in the absence of dedicated computing devices such as GPUs and TPUs. To address these challenges, we propose Pan-LUT, a novel learnable look-up table (LUT) framework for pan-sharpening that strikes a balance between performance and computational efficiency for high-resolution remote sensing images. To finely control the spectral transformation, we devise the PAN-guided look-up table (PGLUT) for channel-wise spectral mapping. To effectively capture fine-grained spatial details and adaptively learn local contexts, we introduce the spatial details look-up table (SDLUT) and adaptive aggregation look-up table (AALUT). Our proposed method contains fewer than 300K parameters and processes a 8K resolution image in under 1 ms using a single NVIDIA GeForce RTX 2080 Ti GPU, demonstrating significantly faster performance compared to other methods. Experiments reveal that Pan-LUT efficiently processes large remote sensing images in a lightweight manner, bridging the gap to real-world applications. Furthermore, our model surpasses SOTA methods in full-resolution scenes under real-world conditions, highlighting its effectiveness and efficiency.
Despite its significant achievements in large-scale scene reconstruction, 3D Gaussian Splatting still faces substantial challenges, including slow processing, high computational costs, and limited geometric accuracy. These core issues arise from its inherently unstructured design and the absence of efficient parallelization. To overcome these challenges simultaneously, we introduce CityGS-X, a scalable architecture built on a novel parallelized hybrid hierarchical 3D representation (PH^2-3D). As an early attempt, CityGS-X abandons the cumbersome merge-and-partition process and instead adopts a newly-designed batch-level multi-task rendering process. This architecture enables efficient multi-GPU rendering through dynamic Level-of-Detail voxel allocations, significantly improving scalability and performance. Through extensive experiments, CityGS-X consistently outperforms existing methods in terms of faster training times, larger rendering capacities, and more accurate geometric details in large-scale scenes. Notably, CityGS-X can train and render a scene with 5,000+ images in just 5 hours using only 4 * 4090 GPUs, a task that would make other alternative methods encounter Out-Of-Memory (OOM) issues and fail completely. This implies that CityGS-X is far beyond the capacity of other existing methods.
Recently, open-source video diffusion models (VDMs), such as WanX, Magic141 and HunyuanVideo, have been scaled to over 10 billion parameters. These large-scale VDMs have demonstrated significant improvements over smaller-scale VDMs across multiple dimensions, including enhanced visual quality and more natural motion dynamics. However, these models face two major limitations: (1) High inference overhead: Large-scale VDMs require approximately 10 minutes to synthesize a 28-step video on a single H100 GPU. (2) Limited in portrait video synthesis: Models like WanX-I2V and HunyuanVideo-I2V often produce unnatural facial expressions and movements in portrait videos. To address these challenges, we propose MagicDistillation, a novel framework designed to reduce inference overhead while ensuring the generalization of VDMs for portrait video synthesis. Specifically, we primarily use sufficiently high-quality talking video to fine-tune Magic141, which is dedicated to portrait video synthesis. We then employ LoRA to effectively and efficiently fine-tune the fake DiT within the step distillation framework known as distribution matching distillation (DMD). Following this, we apply weak-to-strong (W2S) distribution matching and minimize the discrepancy between the fake data distribution and the ground truth distribution, thereby improving the visual fidelity and motion dynamics of the synthesized videos. Experimental results on portrait video synthesis demonstrate the effectiveness of MagicDistillation, as our method surpasses Euler, LCM, and DMD baselines in both FID/FVD metrics and VBench. Moreover, MagicDistillation, requiring only 4 steps, also outperforms WanX-I2V (14B) and HunyuanVideo-I2V (13B) on visualization and VBench. Our project page is https://magicdistillation.github.io/MagicDistillation/.
We propose a novel approach that adapts hierarchical vision foundation models for real-time ultrasound image segmentation. Existing ultrasound segmentation methods often struggle with adaptability to new tasks, relying on costly manual annotations, while real-time approaches generally fail to match state-of-the-art performance. To overcome these limitations, we introduce an adaptive framework that leverages the vision foundation model Hiera to extract multi-scale features, interleaved with DINOv2 representations to enhance visual expressiveness. These enriched features are then decoded to produce precise and robust segmentation. We conduct extensive evaluations on six public datasets and one in-house dataset, covering both cardiac and thyroid ultrasound segmentation. Experiments show that our approach outperforms state-of-the-art methods across multiple datasets and excels with limited supervision, surpassing nnUNet by over 20\% on average in the 1\% and 10\% data settings. Our method achieves $\sim$77 FPS inference speed with TensorRT on a single GPU, enabling real-time clinical applications.
3D Gaussian splatting (3DGS) is a popular radiance field method, with many application-specific extensions. Most variants rely on the same core algorithm: depth-sorting of Gaussian splats then rasterizing in primitive order. This ensures correct alpha compositing, but can cause rendering artifacts due to built-in approximations. Moreover, for a fixed representation, sorted rendering offers little control over render cost and visual fidelity. For example, and counter-intuitively, rendering a lower-resolution image is not necessarily faster. In this work, we address the above limitations by combining 3D Gaussian splatting with stochastic rasterization. Concretely, we leverage an unbiased Monte Carlo estimator of the volume rendering equation. This removes the need for sorting, and allows for accurate 3D blending of overlapping Gaussians. The number of Monte Carlo samples further imbues 3DGS with a way to trade off computation time and quality. We implement our method using OpenGL shaders, enabling efficient rendering on modern GPU hardware. At a reasonable visual quality, our method renders more than four times faster than sorted rasterization.
By provisioning inference offloading services, edge inference drives the rapid growth of AI applications at the network edge. However, achieving high task throughput with stringent latency requirements remains a significant challenge. To address this issue, we develop a parameter-sharing AI model loading (PartialLoading) framework for multi-user edge inference, which exploits two key insights: 1) the majority of latency arises from loading AI models into server GPU memory, and 2) different AI models can share a significant number of parameters, for which redundant loading should be avoided. Towards this end, we formulate a joint multi-user scheduling and spectrum bandwidth allocation problem to maximize task throughput by exploiting shared parameter blocks across models. The intuition is to judiciously schedule user requests to reuse the shared parameter blocks between consecutively loaded models, thereby reducing model loading time substantially. To facilitate solution finding, we decouple the problem into two sub-problems, i.e., user scheduling and bandwidth allocation, showing that solving them sequentially is equivalent to solving the original problem. Due to the NP-hardness of the problem, we first study an important special case called the "bottom-layer-sharing" case, where AI models share some bottom layers within clusters, and design a dynamic programming-based algorithm to obtain the optimal solution in polynomial time. For the general case, where shared parameter blocks appear at arbitrary positions within AI models, we propose a greedy heuristic to obtain the sub-optimal solution efficiently. Simulation results demonstrate that the proposed framework significantly improves task throughput under deadline constraints compared with user scheduling without exploiting parameter sharing.
The deployment of Machine Learning models at cloud have grown by tech companies. Hardware requirements are higher when these models involve Deep Learning (DL) techniques and the cloud providers' costs may be a barrier. We explore deploying DL models using for experiments the GECToR model, a DL solution for Grammatical Error Correction, across three of the major cloud platforms (AWS, Google Cloud, Azure). We evaluate real-time latency, hardware usage and cost at each cloud provider by 7 execution environments with 10 experiments reproduced. We found that while GPUs excel in performance, they had an average cost 300% higher than solutions without GPU. Our analysis also identifies that processor cache size is crucial for cost-effective CPU deployments, enabling over 50% of cost reduction compared to GPUs. This study demonstrates the feasibility and affordability of cloud-based DL inference solutions without GPUs, benefiting resource-constrained users like startups.