Recent investments in cislunar applications open new frontiers for space missions within highly nonlinear dynamical regimes. In this paper, we propose a method based on Sequential Convex Programming (SCP) to loiter around a given target with impulsive actuation while satisfying path constraints continuously over the finite time-horizon, i.e., independently of the number of nodes in which domain is discretized. Location, timing, magnitude, and direction of a fixed number of impulses are optimized in a model predictive framework, exploiting the exact nonlinear dynamics of non-stationary orbital regimes. The proposed approach is first validated on a relative orbiting problem with respect to a selenocentric near rectilinear halo orbit. The approach is then compared to a formulation with path constraints imposed only at nodes and with mesh refined to ensure complete satisfaction of path constraints over the continuous-time horizon. CPU time per iteration of 400 ms for the refined-mesh approach reduce to 5.5 ms for the proposed approach.
We present PixelBrax, a set of continuous control tasks with pixel observations. We combine the Brax physics engine with a pure JAX renderer, allowing reinforcement learning (RL) experiments to run end-to-end on the GPU. PixelBrax can render observations over thousands of parallel environments and can run two orders of magnitude faster than existing benchmarks that rely on CPU-based rendering. Additionally, PixelBrax supports fully reproducible experiments through its explicit handling of any stochasticity within the environments and supports color and video distractors for benchmarking generalization. We open-source PixelBrax alongside JAX implementations of several RL algorithms at github.com/trevormcinroe/pixelbrax.
The machine learning (ML) training over disparate data sources traditionally involves materialization, which can impose substantial time and space overhead due to data movement and replication. Factorized learning, which leverages direct computation on disparate sources through linear algebra (LA) rewriting, has emerged as a viable alternative to improve computational efficiency. However, the adaptation of factorized learning to leverage the full capabilities of modern LA-friendly hardware like GPUs has been limited, often requiring manual intervention for algorithm compatibility. This paper introduces Ilargi, a novel factorized learning framework that utilizes matrix-represented data integration (DI) metadata to facilitate automatic factorization across CPU and GPU environments without the need for costly relational joins. Ilargi incorporates an ML-based cost estimator to intelligently selects between factorization and materialization based on data properties, algorithm complexity, hardware environments, and their interactions. This strategy ensures up to 8.9x speedups on GPUs and achieves over 20% acceleration in batch ML training workloads, thereby enhancing the practicability of ML training across diverse data integration scenarios and hardware platforms. To our knowledge, this work is the very first effort in GPU-compatible factorized learning.
In this methods article, we provide a flexible but easy-to-use implementation of Direct Coupling Analysis (DCA) based on Boltzmann machine learning, together with a tutorial on how to use it. The package \texttt{adabmDCA 2.0} is available in different programming languages (C++, Julia, Python) usable on different architectures (single-core and multi-core CPU, GPU) using a common front-end interface. In addition to several learning protocols for dense and sparse generative DCA models, it allows to directly address common downstream tasks like residue-residue contact prediction, mutational-effect prediction, scoring of sequence libraries and generation of artificial sequences for sequence design. It is readily applicable to protein and RNA sequence data.
Gravitational wave (GW) detectors, such as LIGO, Virgo, and KAGRA, detect faint signals from distant astrophysical events. However, their high sensitivity also makes them susceptible to background noise, which can obscure these signals. This noise often includes transient artifacts called 'glitches', that can mimic genuine astrophysical signals or mask their true characteristics. In this study, we present DeepExtractor, a deep learning framework that is designed to reconstruct signals and glitches with power exceeding interferometer noise, regardless of their source. We design DeepExtractor to model the inherent noise distribution of GW detectors, following conventional assumptions that the noise is Gaussian and stationary over short time scales. It operates by predicting and subtracting the noise component of the data, retaining only the clean reconstruction of signal or glitch. We focus on applications related to glitches and validate DeepExtractor's effectiveness through three experiments: (1) reconstructing simulated glitches injected into simulated detector noise, (2) comparing its performance with the state-of-the-art BayesWave algorithm, and (3) analyzing real data from the Gravity Spy dataset to demonstrate effective glitch subtraction from LIGO strain data. We further demonstrate its potential by reconstructing three real GW events from LIGO's third observing run, without being trained on GW waveforms. Our proposed model achieves a median mismatch of only 0.9% for simulated glitches, outperforming several deep learning baselines. Additionally, DeepExtractor surpasses BayesWave in glitch recovery, offering a dramatic computational speedup by reconstructing one glitch sample in approximately 0.1 seconds on a CPU, compared to BayesWave's processing time of approximately one hour per glitch.
Parallel solving via cube-and-conquer is a key method for scaling SAT solvers to hard instances. While cube-and-conquer has proven successful for pure SAT problems, notably the Pythagorean triples conjecture, its application to SAT solvers extended with propagators presents unique challenges, as these propagators learn constraints dynamically during the search. We study this problem using SAT Modulo Symmetries (SMS) as our primary test case, where a symmetry-breaking propagator reduces the search space by learning constraints that eliminate isomorphic graphs. Through extensive experimentation comprising over 10,000 CPU hours, we systematically evaluate different cube-and-conquer variants on three well-studied combinatorial problems. Our methodology combines prerun phases to collect learned constraints, various cubing strategies, and parameter tuning via algorithm configuration and LLM-generated design suggestions. The comprehensive empirical evaluation provides new insights into effective cubing strategies for propagator-based SAT solving, with our best method achieving speedups of 2-3x from improved cubing and parameter tuning, providing an additional 1.5-2x improvement on harder instances.
The burgeoning field of foundation models necessitates advanced data processing mechanisms capable of harnessing vast and valuable data with various types used by these models. Nevertheless, the current landscape presents unique challenges that traditional data processing frameworks struggle to handle effectively, particularly in handling the complexity of multimodal data. In response, we present Data-Juicer 2.0, a data processing system backed by 100+ data processing operators spanning text, image, video, and audio modalities, supporting more critical tasks including data analysis, synthesis, annotation, and foundation model post-training. With seamless compatibility and dedicated optimization for popular dataset hubs like Hugging Face and computing engines like Ray, it improves upon its predecessor in terms of usability, efficiency, and programmability. It features an easily accessible user interface layer that supports decoupled Python interactions, RESTful APIs, and conversational commands. It contains a new runtime layer optimized for adaptive execution and management across varying dataset scales, processing demands, and computational environments, while hiding unnecessary system details. Extensive empirical evaluations demonstrate Data-Juicer 2.0's remarkable performance and scalability, highlighting its capability to efficiently process TB-level data with 10k+ CPU cores. The system is publicly available and has been widely adopted in diverse research fields and real-world products such as Alibaba Cloud PAI. We actively maintain it and share insights from practical feedback, with the goal of facilitating research and application of next-generation foundation models.
Mixture-of-Experts (MoE) models, though highly effective for various machine learning tasks, face significant deployment challenges on memory-constrained devices. While GPUs offer fast inference, their limited memory compared to CPUs means not all experts can be stored on the GPU simultaneously, necessitating frequent, costly data transfers from CPU memory, often negating GPU speed advantages. To address this, we present DAOP, an on-device MoE inference engine to optimize parallel GPU-CPU execution. DAOP dynamically allocates experts between CPU and GPU based on per-sequence activation patterns, and selectively pre-calculates predicted experts on CPUs to minimize transfer latency. This approach enables efficient resource utilization across various expert cache ratios while maintaining model accuracy through a novel graceful degradation mechanism. Comprehensive evaluations across various datasets show that DAOP outperforms traditional expert caching and prefetching methods by up to 8.20x and offloading techniques by 1.35x while maintaining accuracy.
The dominance of large decoder-only language models has overshadowed encoder-decoder architectures, despite their fundamental efficiency advantages in sequence processing. For small language models (SLMs) - those with 1 billion parameters or fewer - our systematic analysis across GPU, CPU, and NPU platforms reveals that encoder-decoder architectures achieve 47% lower first-token latency and 4.7x higher throughput compared to decoder-only models on edge devices. These gains may be attributed to encoder-decoder's one-time input processing and efficient separation of understanding and generation phases. We introduce a novel knowledge distillation framework that enables encoder-decoder models to leverage capabilities from large scalable decoder-only teachers while preserving their architectural advantages, achieving up to 6 average performance points improvement across diverse tasks, with significant gains in asymmetric sequence tasks where input and output distributions can benefit from different processing approaches. When combined with modern advances like Rotary Positional Embeddings (RoPE) and Vision encoders, our systematic investigation demonstrates that encoder-decoder architectures provide a more practical path toward deploying capable language models in resource-constrained environments. Our findings challenge the prevailing trend toward decoder-only scaling, showing that architectural choices become increasingly crucial as parameter budgets decrease, particularly for on-device and edge deployments where computational efficiency is paramount.
We present MOFA, an open-source generative AI (GenAI) plus simulation workflow for high-throughput generation of metal-organic frameworks (MOFs) on large-scale high-performance computing (HPC) systems. MOFA addresses key challenges in integrating GPU-accelerated computing for GPU-intensive GenAI tasks, including distributed training and inference, alongside CPU- and GPU-optimized tasks for screening and filtering AI-generated MOFs using molecular dynamics, density functional theory, and Monte Carlo simulations. These heterogeneous tasks are unified within an online learning framework that optimizes the utilization of available CPU and GPU resources across HPC systems. Performance metrics from a 450-node (14,400 AMD Zen 3 CPUs + 1800 NVIDIA A100 GPUs) supercomputer run demonstrate that MOFA achieves high-throughput generation of novel MOF structures, with CO$_2$ adsorption capacities ranking among the top 10 in the hypothetical MOF (hMOF) dataset. Furthermore, the production of high-quality MOFs exhibits a linear relationship with the number of nodes utilized. The modular architecture of MOFA will facilitate its integration into other scientific applications that dynamically combine GenAI with large-scale simulations.
This paper presents a novel approach for similarity search with complex filtering capabilities on billion-scale datasets, optimized for CPU inference. Our method extends the classical IVF-Flat index structure to integrate multi-dimensional filters. The proposed algorithm combines dense embeddings with discrete filtering attributes, enabling fast retrieval in high-dimensional spaces. Designed specifically for CPU-based systems, our disk-based approach offers a cost-effective solution for large-scale similarity search. We demonstrate the effectiveness of our method through a case study, showcasing its potential for various practical uses.
Recognizing motions of tiny targets - only few dozen pixels - in cluttered backgrounds remains a fundamental challenge when standard feature-based or deep learning methods fail under scarce visual cues. We propose STMDNet, a model-based computational framework to Recognize motions of tiny targets at variable velocities under low-sampling frequency scenarios. STMDNet designs a novel dual-dynamics-and-correlation mechanism, harnessing ipsilateral excitation to integrate target cues and leakage-enhancing-type contralateral inhibition to suppress large-object and background motion interference. Moreover, we develop the first collaborative directional encoding-decoding strategy that determines the motion direction from only one correlation per spatial location, cutting computational costs to one-eighth of prior methods. Further, simply substituting the backbone of a strong STMD model with STMDNet raises AUC by 24%, yielding an enhanced STMDNet-F. Evaluations on real-world low sampling frequency datasets show state-of-the-art results, surpassing the deep learning baseline. Across diverse speeds, STMDNet-F improves mF1 by 19%, 16%, and 8% at 240Hz, 120Hz, and 60Hz, respectively, while STMDNet achieves 87 FPS on a single CPU thread. These advances highlight STMDNet as a next-generation backbone for tiny target motion pattern recognition and underscore its broader potential to revitalize model-based visual approaches in motion detection.
This paper introduces HEPPO, an FPGA-based accelerator designed to optimize the Generalized Advantage Estimation (GAE) stage in Proximal Policy Optimization (PPO). Unlike previous approaches that focused on trajectory collection and actor-critic updates, HEPPO addresses GAE's computational demands with a parallel, pipelined architecture implemented on a single System-on-Chip (SoC). This design allows for the adaptation of various hardware accelerators tailored for different PPO phases. A key innovation is our strategic standardization technique, which combines dynamic reward standardization and block standardization for values, followed by 8-bit uniform quantization. This method stabilizes learning, enhances performance, and manages memory bottlenecks, achieving a 4x reduction in memory usage and a 1.5x increase in cumulative rewards. We propose a solution on a single SoC device with programmable logic and embedded processors, delivering throughput orders of magnitude higher than traditional CPU-GPU systems. Our single-chip solution minimizes communication latency and throughput bottlenecks, significantly boosting PPO training efficiency. Experimental results show a 30% increase in PPO speed and a substantial reduction in memory access time, underscoring HEPPO's potential for broad applicability in hardware-efficient reinforcement learning algorithms.
While ML model training and inference are both GPU-intensive, CPU-based data processing is often the bottleneck. Distributed data processing systems based on the batch or stream processing models assume homogeneous resource requirements. They excel at CPU-based computation but either under-utilize heterogeneous resources or impose high overheads on failure and reconfiguration. We introduce the streaming batch model, a hybrid of the two models that enables efficient and fault-tolerant heterogeneous execution. The key idea is to execute one partition at a time to allow lineage-based recovery with dynamic resource allocation. This enables memory-efficient pipelining across heterogeneous resources, similar to stream processing, but also offers the elasticity and fault tolerance properties of batch processing. We present Ray Data, an implementation of the streaming batch model that improves throughput on heterogeneous batch inference pipelines by 3--8$\times$ compared to traditional batch and stream processing systems. When training Stable Diffusion, Ray Data matches the throughput of single-node ML data loaders while additionally leveraging distributed heterogeneous clusters to further improve training throughput by 31%.
Keeping ML-based recommender models up-to-date as data drifts and evolves is essential to maintain accuracy. As a result, online data preprocessing plays an increasingly important role in serving recommender systems. Existing solutions employ multiple CPU workers to saturate the input bandwidth of a single training node. Such an approach results in high deployment costs and energy consumption. For instance, a recent report from industrial deployments shows that data storage and ingestion pipelines can account for over 60\% of the power consumption in a recommender system. In this paper, we tackle the issue from a hardware perspective by introducing Piper, a flexible and network-attached accelerator that executes data loading and preprocessing pipelines in a streaming fashion. As part of the design, we define MiniPipe, the smallest pipeline unit enabling multi-pipeline implementation by executing various data preprocessing tasks across the single board, giving Piper the ability to be reconfigured at runtime. Our results, using publicly released commercial pipelines, show that Piper, prototyped on a power-efficient FPGA, achieves a 39$\sim$105$\times$ speedup over a server-grade, 128-core CPU and 3$\sim$17$\times$ speedup over GPUs like RTX 3090 and A100 in multiple pipelines. The experimental analysis demonstrates that Piper provides advantages in both latency and energy efficiency for preprocessing tasks in recommender systems, providing an alternative design point for systems that today are in very high demand.
We introduce Glinthawk, an architecture for offline Large Language Model (LLM) inference. By leveraging a two-tiered structure, Glinthawk optimizes the utilization of the high-end accelerators ("Tier 1") by offloading the attention mechanism to lower-end compute tier ("Tier 2"). This separation allows the memory demand of the attention, known as the key-value cache, to scale independently from the model weights, enabling larger batch sizes and more efficient accelerator usage. Prototyped with NVIDIA T4 GPUs and standard CPU VMs, Glinthawk improves throughput by $5.9\times$ and reduces cost of generation by $2.8\times$, compared to paged attention baselines. For long sequence lengths, it achieves $16.3\times$ throughput improvement at $2.4\times$ less cost. Our evaluation shows that this architecture can tolerate moderate network latency with minimal performance degradation, making it highly effective for latency-tolerant, throughput-focused applications such as batch processing. The prototype is publicly available at https://github.com/microsoft/glinthawk.
Gaussian Mixture Models (GMMs) range among the most frequently used machine learning models. However, training large, general GMMs becomes computationally prohibitive for datasets with many data points $N$ of high-dimensionality $D$. For GMMs with arbitrary covariances, we here derive a highly efficient variational approximation, which is integrated with mixtures of factor analyzers (MFAs). For GMMs with $C$ components, our proposed algorithm significantly reduces runtime complexity per iteration from $\mathcal{O}(NCD^2)$ to a complexity scaling linearly with $D$ and remaining constant w.r.t. $C$. Numerical validation of this theoretical complexity reduction then shows the following: the distance evaluations required for the entire GMM optimization process scale sublinearly with $NC$. On large-scale benchmarks, this sublinearity results in speed-ups of an order-of-magnitude compared to the state-of-the-art. As a proof of concept, we train GMMs with over 10 billion parameters on about 100 million images, and observe training times of approximately nine hours on a single state-of-the-art CPU.
3D models surpass 2D models in CT/MRI segmentation by effectively capturing inter-slice relationships. However, the added depth dimension substantially increases memory consumption. While patch-based training alleviates memory constraints, it significantly slows down the inference speed due to the sliding window (SW) approach. We propose No-More-Sliding-Window (NMSW), a novel end-to-end trainable framework that enhances the efficiency of generic 3D segmentation backbone during an inference step by eliminating the need for SW. NMSW employs a differentiable Top-k module to selectively sample only the most relevant patches, thereby minimizing redundant computations. When patch-level predictions are insufficient, the framework intelligently leverages coarse global predictions to refine results. Evaluated across 3 tasks using 3 segmentation backbones, NMSW achieves competitive accuracy compared to SW inference while significantly reducing computational complexity by 91% (88.0 to 8.00 TMACs). Moreover, it delivers a 9.1x faster inference on the H100 GPU (99.0 to 8.3 sec) and a 11.1x faster inference on the Xeon Gold CPU (2110 to 189 sec). NMSW is model-agnostic, further boosting efficiency when integrated with any existing efficient segmentation backbones.
Zero-shot Text-To-Speech (TTS) synthesis shows great promise for personalized voice customization through voice cloning. However, current methods for achieving zero-shot TTS heavily rely on large model scales and extensive training datasets to ensure satisfactory performance and generalizability across various speakers. This raises concerns regarding both deployment costs and data security. In this paper, we present a lightweight and stable zero-shot TTS system. We introduce a novel TTS architecture designed to effectively model linguistic content and various speaker attributes from source speech and prompt speech, respectively. Furthermore, we present a two-stage self-distillation framework that constructs parallel data pairs for effectively disentangling linguistic content and speakers from the perspective of training data. Extensive experiments show that our system exhibits excellent performance and superior stability on the zero-shot TTS tasks. Moreover, it shows markedly superior computational efficiency, with RTFs of 0.13 and 0.012 on the CPU and GPU, respectively.
In this paper we introduce Keras Sig a high-performance pythonic library designed to compute path signature for deep learning applications. Entirely built in Keras 3, \textit{Keras Sig} leverages the seamless integration with the mostly used deep learning backends such as PyTorch, JAX and TensorFlow. Inspired by Kidger and Lyons (2021),we proposed a novel approach reshaping signature calculations to leverage GPU parallelism. This adjustment allows us to reduce the training time by 55\% and 5 to 10-fold improvements in direct signature computation compared to existing methods, while maintaining similar CPU performance. Relying on high-level tensor operations instead of low-level C++ code, Keras Sig significantly reduces the versioning and compatibility issues commonly encountered in deep learning libraries, while delivering superior or comparable performance across various hardware configurations. We demonstrate through extensive benchmarking that our approach scales efficiently with the length of input sequences and maintains competitive performance across various signature parameters, though bounded by memory constraints for very large signature dimensions.