Integrating hyperscale AI into national defense M&S(Modeling and Simulation), under the expanding IoMDT(Internet of Military Defense Things) framework, is crucial for boosting strategic and operational readiness. We examine how IoMDT-driven hyperscale AI can provide high accuracy, speed, and the ability to simulate complex, interconnected battlefield scenarios in defense M&S. Countries like the United States and China are leading the adoption of these technologies, with varying levels of success. However, realizing the full potential of hyperscale AI requires overcoming challenges such as closed networks, sparse or long-tail data, complex decision-making processes, and a shortage of experts. Future directions highlight the need to adopt domestic foundation models, expand GPU/NPU investments, leverage large tech services, and employ open source solutions. These efforts will enhance national security, maintain a competitive edge, and spur broader technological and economic growth. With this blueprint, the Republic of Korea can strengthen its defense posture and stay ahead of emerging threats in modern warfare.
The computational and memory demands of vanilla attention scale quadratically with the sequence length $N$, posing significant challenges for processing long sequences in Transformer models. FlashAttention alleviates these challenges by eliminating the $O(N^2)$ memory dependency and reducing attention latency through IO-aware memory optimizations. However, its native support for certain attention mask types is limited, and it does not inherently accommodate more complex masking requirements. Previous approaches resort to using dense masks with $O(N^2)$ memory complexity, leading to inefficiencies. In this paper, we propose FlashMask, an extension of FlashAttention that introduces a column-wise sparse representation of attention masks. This approach efficiently represents a wide range of mask types and facilitates the development of optimized kernel implementations. By adopting this novel representation, FlashMask achieves linear memory complexity $O(N)$, suitable for modeling long-context sequences. Moreover, this representation enables kernel optimizations that eliminate unnecessary computations by leveraging sparsity in the attention mask, without sacrificing computational accuracy, resulting in higher computational efficiency. We evaluate FlashMask's performance in fine-tuning and alignment training of LLMs such as SFT, LoRA, DPO, and RM. FlashMask achieves significant throughput improvements, with end-to-end speedups ranging from 1.65x to 3.22x compared to existing FlashAttention dense method. Additionally, our kernel-level comparisons demonstrate that FlashMask surpasses the latest counterpart, FlexAttention, by 12.1% to 60.7% in terms of kernel TFLOPs/s, achieving 37.8% to 62.3% of the theoretical maximum FLOPs/s on the A100 GPU. The code is open-sourced on PaddlePaddle and integrated into PaddleNLP, supporting models with over 100 billion parameters for contexts up to 128K tokens.
Many applications are leveraging large language models (LLMs) for complex tasks, and they generally demand low inference latency and high serving throughput for interactive online jobs such as chatbots. However, the tight latency requirement and high load variance of applications pose challenges to serving systems in achieving high GPU utilization. Due to the high costs of scheduling and preemption, today's systems generally use separate clusters to serve online and offline inference tasks, and dedicate GPUs for online inferences to avoid interference. This approach leads to underutilized GPUs because one must reserve enough GPU resources for the peak expected load, even if the average load is low. This paper proposes to harvest stranded GPU resources for offline LLM inference tasks such as document summarization and LLM benchmarking. Unlike online inferences, these tasks usually run in a batch-processing manner with loose latency requirements, making them a good fit for stranded resources that are only available shortly. To enable safe and efficient GPU harvesting without interfering with online tasks, we built ConServe, an LLM serving system that contains (1) an execution engine that preempts running offline tasks upon the arrival of online tasks, (2) an incremental checkpointing mechanism that minimizes the amount of recomputation required by preemptions, and (3) a scheduler that adaptively batches offline tasks for higher GPU utilization. Our evaluation demonstrates that ConServe achieves strong performance isolation when co-serving online and offline tasks but at a much higher GPU utilization. When colocating practical online and offline workloads on popular models such as Llama-2-7B, ConServe achieves 2.35$\times$ higher throughput than state-of-the-art online serving systems and reduces serving latency by 84$\times$ compared to existing co-serving systems.
Symbolic regression (SR) is a powerful machine learning approach that searches for both the structure and parameters of algebraic models, offering interpretable and compact representations of complex data. Unlike traditional regression methods, SR explores progressively complex feature spaces, which can uncover simple models that generalize well, even from small datasets. Among SR algorithms, the Sure Independence Screening and Sparsifying Operator (SISSO) has proven particularly effective in the natural sciences, helping to rediscover fundamental physical laws as well as discover new interpretable equations for materials property modeling. However, its widespread adoption has been limited by performance inefficiencies and the challenges posed by its FORTRAN-based implementation, especially in modern computing environments. In this work, we introduce TorchSISSO, a native Python implementation built in the PyTorch framework. TorchSISSO leverages GPU acceleration, easy integration, and extensibility, offering a significant speed-up and improved accuracy over the original. We demonstrate that TorchSISSO matches or exceeds the performance of the original SISSO across a range of tasks, while dramatically reducing computational time and improving accessibility for broader scientific applications.
This work develops a distributed graph neural network (GNN) methodology for mesh-based modeling applications using a consistent neural message passing layer. As the name implies, the focus is on enabling scalable operations that satisfy physical consistency via halo nodes at sub-graph boundaries. Here, consistency refers to the fact that a GNN trained and evaluated on one rank (one large graph) is arithmetically equivalent to evaluations on multiple ranks (a partitioned graph). This concept is demonstrated by interfacing GNNs with NekRS, a GPU-capable exascale CFD solver developed at Argonne National Laboratory. It is shown how the NekRS mesh partitioning can be linked to the distributed GNN training and inference routines, resulting in a scalable mesh-based data-driven modeling workflow. We study the impact of consistency on the scalability of mesh-based GNNs, demonstrating efficient scaling in consistent GNNs for up to O(1B) graph nodes on the Frontier exascale supercomputer.
This paper introduces a new approach to extract and analyze vector data from technical drawings in PDF format. Our method involves converting PDF files into SVG format and creating a feature-rich graph representation, which captures the relationships between vector entities using geometrical information. We then apply a graph attention transformer with hierarchical label definition to achieve accurate line-level segmentation. Our approach is evaluated on two datasets, including the public FloorplanCAD dataset, which achieves state-of-the-art results on weighted F1 score, surpassing existing methods. The proposed vector-based method offers a more scalable solution for large-scale technical drawing analysis compared to vision-based approaches, while also requiring significantly less GPU power than current state-of-the-art vector-based techniques. Moreover, it demonstrates improved performance in terms of the weighted F1 (wF1) score on the semantic segmentation task. Our results demonstrate the effectiveness of our approach in extracting meaningful information from technical drawings, enabling new applications, and improving existing workflows in the AEC industry. Potential applications of our approach include automated building information modeling (BIM) and construction planning, which could significantly impact the efficiency and productivity of the industry.
Traditional end-to-end deep learning models often enhance feature representation and overall performance by increasing the depth and complexity of the network during training. However, this approach inevitably introduces issues of parameter redundancy and resource inefficiency, especially in deeper networks. While existing works attempt to skip certain redundant layers to alleviate these problems, challenges related to poor performance, computational complexity, and inefficient memory usage remain. To address these issues, we propose an innovative training approach called Replacement Learning, which mitigates these limitations by completely replacing all the parameters of the frozen layers with only two learnable parameters. Specifically, Replacement Learning selectively freezes the parameters of certain layers, and the frozen layers utilize parameters from adjacent layers, updating them through a parameter integration mechanism controlled by two learnable parameters. This method leverages information from surrounding structures, reduces computation, conserves GPU memory, and maintains a balance between historical context and new inputs, ultimately enhancing overall model performance. We conducted experiments across four benchmark datasets, including CIFAR-10, STL-10, SVHN, and ImageNet, utilizing various architectures such as CNNs and ViTs to validate the effectiveness of Replacement Learning. Experimental results demonstrate that our approach reduces the number of parameters, training time, and memory consumption while completely surpassing the performance of end-to-end training.
The expanding context windows in large language models (LLMs) have greatly enhanced their capabilities in various applications, but they also introduce significant challenges in maintaining low latency, particularly in Time to First Token (TTFT). This paper identifies that the sharp rise in TTFT as context length increases is predominantly driven by queuing delays, which are caused by the growing demands for GPU Key-Value (KV) cache allocation clashing with the limited availability of KV cache blocks. To address this issue, we propose LayerKV, a simple yet effective plug-in method that effectively reduces TTFT without requiring additional hardware or compromising output performance, while seamlessly integrating with existing parallelism strategies and scheduling techniques. Specifically, LayerKV introduces layer-wise KV block allocation, management, and offloading for fine-grained control over system memory, coupled with an SLO-aware scheduler to optimize overall Service Level Objectives (SLOs). Comprehensive evaluations on representative models, ranging from 7B to 70B parameters, across various GPU configurations, demonstrate that LayerKV improves TTFT latency up to 69x and reduces SLO violation rates by 28.7%, significantly enhancing the user experience.
Traffic forecasting is a cornerstone of smart city management, enabling efficient resource allocation and transportation planning. Deep learning, with its ability to capture complex nonlinear patterns in spatiotemporal (ST) data, has emerged as a powerful tool for traffic forecasting. While graph neural networks (GCNs) and transformer-based models have shown promise, their computational demands often hinder their application to real-world road networks, particularly those with large-scale spatiotemporal interactions. To address these challenges, we propose a novel spatiotemporal graph transformer (STGformer) architecture. STGformer effectively balances the strengths of GCNs and Transformers, enabling efficient modeling of both global and local traffic patterns while maintaining a manageable computational footprint. Unlike traditional approaches that require multiple attention layers, STG attention block captures high-order spatiotemporal interactions in a single layer, significantly reducing computational cost. In particular, STGformer achieves a 100x speedup and a 99.8\% reduction in GPU memory usage compared to STAEformer during batch inference on a California road graph with 8,600 sensors. We evaluate STGformer on the LargeST benchmark and demonstrate its superiority over state-of-the-art Transformer-based methods such as PDFormer and STAEformer, which underline STGformer's potential to revolutionize traffic forecasting by overcoming the computational and memory limitations of existing approaches, making it a promising foundation for future spatiotemporal modeling tasks.
Simulation has enabled unprecedented compute-scalable approaches to robot learning. However, many existing simulation frameworks typically support a narrow range of scenes/tasks and lack features critical for scaling generalizable robotics and sim2real. We introduce and open source ManiSkill3, the fastest state-visual GPU parallelized robotics simulator with contact-rich physics targeting generalizable manipulation. ManiSkill3 supports GPU parallelization of many aspects including simulation+rendering, heterogeneous simulation, pointclouds/voxels visual input, and more. Simulation with rendering on ManiSkill3 can run 10-1000x faster with 2-3x less GPU memory usage than other platforms, achieving up to 30,000+ FPS in benchmarked environments due to minimal python/pytorch overhead in the system, simulation on the GPU, and the use of the SAPIEN parallel rendering system. Tasks that used to take hours to train can now take minutes. We further provide the most comprehensive range of GPU parallelized environments/tasks spanning 12 distinct domains including but not limited to mobile manipulation for tasks such as drawing, humanoids, and dextrous manipulation in realistic scenes designed by artists or real-world digital twins. In addition, millions of demonstration frames are provided from motion planning, RL, and teleoperation. ManiSkill3 also provides a comprehensive set of baselines that span popular RL and learning-from-demonstrations algorithms.
Neural simulation-based inference (SBI) describes an emerging family of methods for Bayesian inference with intractable likelihood functions that use neural networks as surrogate models. Here we introduce sbijax, a Python package that implements a wide variety of state-of-the-art methods in neural simulation-based inference using a user-friendly programming interface. sbijax offers high-level functionality to quickly construct SBI estimators, and compute and visualize posterior distributions with only a few lines of code. In addition, the package provides functionality for conventional approximate Bayesian computation, to compute model diagnostics, and to automatically estimate summary statistics. By virtue of being entirely written in JAX, sbijax is extremely computationally efficient, allowing rapid training of neural networks and executing code automatically in parallel on both CPU and GPU.
Diffusion-based generative models have demonstrated their powerful performance across various tasks, but this comes at a cost of the slow sampling speed. To achieve both efficient and high-quality synthesis, various distillation-based accelerated sampling methods have been developed recently. However, they generally require time-consuming fine tuning with elaborate designs to achieve satisfactory performance in a specific number of function evaluation (NFE), making them difficult to employ in practice. To address this issue, we propose Simple and Fast Distillation (SFD) of diffusion models, which simplifies the paradigm used in existing methods and largely shortens their fine-tuning time up to 1000$\times$. We begin with a vanilla distillation-based sampling method and boost its performance to state of the art by identifying and addressing several small yet vital factors affecting the synthesis efficiency and quality. Our method can also achieve sampling with variable NFEs using a single distilled model. Extensive experiments demonstrate that SFD strikes a good balance between the sample quality and fine-tuning costs in few-step image generation task. For example, SFD achieves 4.53 FID (NFE=2) on CIFAR-10 with only 0.64 hours of fine-tuning on a single NVIDIA A100 GPU. Our code is available at https://github.com/zju-pi/diff-sampler.
In this paper, we introduce HEADS-UP, the first egocentric dataset collected from head-mounted cameras, designed specifically for trajectory prediction in blind assistance systems. With the growing population of blind and visually impaired individuals, the need for intelligent assistive tools that provide real-time warnings about potential collisions with dynamic obstacles is becoming critical. These systems rely on algorithms capable of predicting the trajectories of moving objects, such as pedestrians, to issue timely hazard alerts. However, existing datasets fail to capture the necessary information from the perspective of a blind individual. To address this gap, HEADS-UP offers a novel dataset focused on trajectory prediction in this context. Leveraging this dataset, we propose a semi-local trajectory prediction approach to assess collision risks between blind individuals and pedestrians in dynamic environments. Unlike conventional methods that separately predict the trajectories of both the blind individual (ego agent) and pedestrians, our approach operates within a semi-local coordinate system, a rotated version of the camera's coordinate system, facilitating the prediction process. We validate our method on the HEADS-UP dataset and implement the proposed solution in ROS, performing real-time tests on an NVIDIA Jetson GPU through a user study. Results from both dataset evaluations and live tests demonstrate the robustness and efficiency of our approach.
Transformer networks, driven by self-attention, are central to Large Language Models. In generative Transformers, self-attention uses cache memory to store token projections, avoiding recomputation at each time step. However, GPU-stored projections must be loaded into SRAM for each new generation step, causing latency and energy bottlenecks. We present a custom self-attention in-memory computing architecture based on emerging charge-based memories called gain cells, which can be efficiently written to store new tokens during sequence generation and enable parallel analog dot-product computation required for self-attention. However, the analog gain cell circuits introduce non-idealities and constraints preventing the direct mapping of pre-trained models. To circumvent this problem, we design an initialization algorithm achieving text processing performance comparable to GPT-2 without training from scratch. Our architecture respectively reduces attention latency and energy consumption by up to two and five orders of magnitude compared to GPUs, marking a significant step toward ultra-fast, low-power generative Transformers.
Training deep learning models is a repetitive and resource-intensive process. Data scientists often train several models before landing on set of parameters (e.g., hyper-parameter tuning), model architecture (e.g., neural architecture search), among other things that yields the highest accuracy. The computational efficiency of these training tasks depends highly on how well we can supply the training process with training data. The repetitive nature of these tasks results in the same data processing pipelines running over and over exacerbating the need for and costs of computational resources. In this paper, we present Tensorsocket to reduce the computational needs of deep learning training by enabling simultaneous training processes to share the same data loader. Tensorsocket mitigates CPU-side bottlenecks in cases where the collocated training workloads have high throughput on GPU, but are held back by lower data-loading throughput on CPU. Tensorsocket achieves this by reducing redundant computations across collocated training processes and leveraging modern GPU-GPU interconnects. We demonstrate the hardware- and pipeline-agnostic nature of Tensorsocket and evaluate it using a variety of training scenarios. Our evaluation shows that Tensorsocket enables scenarios that are infeasible without data sharing, increases training throughput by up to $100\%$, and when utilizing cloud instances, Tensorsocket achieves cost savings of $50\%$ by reducing the hardware resource needs on the CPU side. Furthermore, Tensorsocket outperforms the state-of-the-art solutions for shared data loading such as CoorDL and Joader. It is easier to use, maintain, and deploy, and either achieves higher or matches the throughput of other solutions while requiring less CPU resources.
Scalability issue plays a crucial role in productionizing modern recommender systems. Even lightweight architectures may suffer from high computational overload due to intermediate calculations, limiting their practicality in real-world applications. Specifically, applying full Cross-Entropy (CE) loss often yields state-of-the-art performance in terms of recommendations quality. Still, it suffers from excessive GPU memory utilization when dealing with large item catalogs. This paper introduces a novel Scalable Cross-Entropy (SCE) loss function in the sequential learning setup. It approximates the CE loss for datasets with large-size catalogs, enhancing both time efficiency and memory usage without compromising recommendations quality. Unlike traditional negative sampling methods, our approach utilizes a selective GPU-efficient computation strategy, focusing on the most informative elements of the catalog, particularly those most likely to be false positives. This is achieved by approximating the softmax distribution over a subset of the model outputs through the maximum inner product search. Experimental results on multiple datasets demonstrate the effectiveness of SCE in reducing peak memory usage by a factor of up to 100 compared to the alternatives, retaining or even exceeding their metrics values. The proposed approach also opens new perspectives for large-scale developments in different domains, such as large language models.
In latent diffusion models (LDMs), denoising diffusion process efficiently takes place on latent space whose dimension is lower than that of pixel space. Decoder is typically used to transform the representation in latent space to that in pixel space. While a decoder is assumed to have an encoder as an accurate inverse, exact encoder-decoder pair rarely exists in practice even though applications often require precise inversion of decoder. Prior works for decoder inversion in LDMs employed gradient descent inspired by inversions of generative adversarial networks. However, gradient-based methods require larger GPU memory and longer computation time for larger latent space. For example, recent video LDMs can generate more than 16 frames, but GPUs with 24 GB memory can only perform gradient-based decoder inversion for 4 frames. Here, we propose an efficient gradient-free decoder inversion for LDMs, which can be applied to diverse latent models. Theoretical convergence property of our proposed inversion has been investigated not only for the forward step method, but also for the inertial Krasnoselskii-Mann (KM) iterations under mild assumption on cocoercivity that is satisfied by recent LDMs. Our proposed gradient-free method with Adam optimizer and learning rate scheduling significantly reduced computation time and memory usage over prior gradient-based methods and enabled efficient computation in applications such as noise-space watermarking while achieving comparable error levels.
Unsupervised domain adaptation (UDA) is essential for medical image segmentation, especially in cross-modality data scenarios. UDA aims to transfer knowledge from a labeled source domain to an unlabeled target domain, thereby reducing the dependency on extensive manual annotations. This paper presents DRL-STNet, a novel framework for cross-modality medical image segmentation that leverages generative adversarial networks (GANs), disentangled representation learning (DRL), and self-training (ST). Our method leverages DRL within a GAN to translate images from the source to the target modality. Then, the segmentation model is initially trained with these translated images and corresponding source labels and then fine-tuned iteratively using a combination of synthetic and real images with pseudo-labels and real labels. The proposed framework exhibits superior performance in abdominal organ segmentation on the FLARE challenge dataset, surpassing state-of-the-art methods by 11.4% in the Dice similarity coefficient and by 13.1% in the Normalized Surface Dice metric, achieving scores of 74.21% and 80.69%, respectively. The average running time is 41 seconds, and the area under the GPU memory-time curve is 11,292 MB. These results indicate the potential of DRL-STNet for enhancing cross-modality medical image segmentation tasks.
GPUs are known to be power-hungry, and due to the boom in artificial intelligence, they are currently the major contributors to the high power demands of upcoming datacenters. Most GPU usage in these popular workloads consist of large general matrix-matrix multiplications (GEMMs), which have therefore been optimized to achieve high utilization of hardware resources. In this work, we show that modifying the input data to GEMMs, while maintaining the matrix shapes and sizes can notably change the power consumption of these kernels. We experiment with four kinds of input variations: value distribution, bit similarity, placement, and sparsity, across different data types. Our findings indicate that these variations can change the GPU power usage during GEMM by almost 40%. We hypothesize that input-dependent power usage variations occur due to changes in the number of bit flips in the GPUs. We propose leveraging this property through compiler and scheduler optimizations to manage power and reduce energy consumption.
In this paper, we introduce Behavior4All, a comprehensive, open-source toolkit for in-the-wild facial behavior analysis, integrating Face Localization, Valence-Arousal Estimation, Basic Expression Recognition and Action Unit Detection, all within a single framework. Available in both CPU-only and GPU-accelerated versions, Behavior4All leverages 12 large-scale, in-the-wild datasets consisting of over 5 million images from diverse demographic groups. It introduces a novel framework that leverages distribution matching and label co-annotation to address tasks with non-overlapping annotations, encoding prior knowledge of their relatedness. In the largest study of its kind, Behavior4All outperforms both state-of-the-art and toolkits in overall performance as well as fairness across all databases and tasks. It also demonstrates superior generalizability on unseen databases and on compound expression recognition. Finally, Behavior4All is way times faster than other toolkits.