Background: As empirical software engineering evolves, more studies adopt data strategies$-$approaches that investigate digital artifacts such as models, source code, or system logs rather than relying on human subjects. Synthesizing results from such studies introduces new methodological challenges. Aims: This study assesses the effects of model quantization on correctness and resource efficiency in deep learning (DL) systems. Additionally, it explores the methodological implications of aggregating evidence from empirical studies that adopt data strategies. Method: We conducted a research synthesis of six primary studies that empirically evaluate model quantization. We applied the Structured Synthesis Method (SSM) to aggregate the findings, which combines qualitative and quantitative evidence through diagrammatic modeling. A total of 19 evidence models were extracted and aggregated. Results: The aggregated evidence indicates that model quantization weakly negatively affects correctness metrics while consistently improving resource efficiency metrics, including storage size, inference latency, and GPU energy consumption$-$a manageable trade-off for many DL deployment contexts. Evidence across quantization techniques remains fragmented, underscoring the need for more focused empirical studies per technique. Conclusions: Model quantization offers substantial efficiency benefits with minor trade-offs in correctness, making it a suitable optimization strategy for resource-constrained environments. This study also demonstrates the feasibility of using SSM to synthesize findings from data strategy-based research.
In this paper, we present a 3D reconstruction and rendering framework termed Mesh-Learner that is natively compatible with traditional rasterization pipelines. It integrates mesh and spherical harmonic (SH) texture (i.e., texture filled with SH coefficients) into the learning process to learn each mesh s view-dependent radiance end-to-end. Images are rendered by interpolating surrounding SH Texels at each pixel s sampling point using a novel interpolation method. Conversely, gradients from each pixel are back-propagated to the related SH Texels in SH textures. Mesh-Learner exploits graphic features of rasterization pipeline (texture sampling, deferred rendering) to render, which makes Mesh-Learner naturally compatible with tools (e.g., Blender) and tasks (e.g., 3D reconstruction, scene rendering, reinforcement learning for robotics) that are based on rasterization pipelines. Our system can train vast, unlimited scenes because we transfer only the SH textures within the frustum to the GPU for training. At other times, the SH textures are stored in CPU RAM, which results in moderate GPU memory usage. The rendering results on interpolation and extrapolation sequences in the Replica and FAST-LIVO2 datasets achieve state-of-the-art performance compared to existing state-of-the-art methods (e.g., 3D Gaussian Splatting and M2-Mapping). To benefit the society, the code will be available at https://github.com/hku-mars/Mesh-Learner.
Domain-adaptive continual pretraining (DAPT) is a state-of-the-art technique that further trains a language model (LM) on its pretraining task, e.g., language masking. Although popular, it requires a significant corpus of domain-related data, which is difficult to obtain for specific domains in languages other than English, such as the process industry in the German language. This paper introduces an efficient approach called ICL-augmented pretraining or ICL-APT that leverages in-context learning (ICL) and k-nearest neighbors (kNN) to augment target data with domain-related and in-domain texts, significantly reducing GPU time while maintaining strong model performance. Our results show that this approach performs better than traditional DAPT by 3.5 points of the average IR metrics (e.g., mAP, MRR, and nDCG) and requires almost 4 times less computing time, providing a cost-effective solution for industries with limited computational capacity. The findings highlight the broader applicability of this framework to other low-resource industries, making NLP-based solutions more accessible and feasible in production environments.
Deep learning (DL) workloads mainly run on accelerators like GPUs. Recent DL quantization techniques demand a new matrix multiplication operator with mixed input data types, further complicating GPU optimization. Prior high-level compilers like Triton lack the expressiveness to implement key optimizations like fine-grained data pipelines and hardware-friendly memory layouts for these operators, while low-level programming models, such as Hidet, Graphene, and CUTLASS, require significant programming efforts. To balance expressiveness with engineering effort, we propose Hexcute, a tile-based programming language that exposes shared memory and register abstractions to enable fine-grained optimization for these operators. Additionally, Hexcute leverages task mapping to schedule the GPU program, and to reduce programming efforts, it automates layout and task mapping synthesis with a novel type-inference-based algorithm. Our evaluation shows that Hexcute generalizes to a wide range of DL operators, achieves 1.7-11.28$\times$ speedup over existing DL compilers for mixed-type operators, and brings up to 2.91$\times$ speedup in the end-to-end evaluation.
The surging demand for GPUs in datacenters for machine learning (ML) has made efficient GPU utilization crucial. However, meeting the diverse needs of ML models while optimizing resource usage is challenging. To enable transparent, fine-grained GPU management that maximizes utilization and energy efficiency while maintaining strong isolation, an operating system (OS) approach is needed. This paper introduces LithOS, a first step toward a GPU OS. LithOS includes the following new abstractions and mechanisms for efficient GPU resource management: (i) a novel TPC Scheduler that supports spatial scheduling at the granularity of individual TPCs, unlocking efficient TPC stealing between workloads; (ii) transparent kernel atomization to reduce head-of-line blocking and enable dynamic resource reallocation mid-execution; (iii) a lightweight hardware right-sizing mechanism that determines the minimal TPC resources needed per atom; and (iv) a transparent power management mechanism that reduces power consumption based on in-flight work behavior. We implement LithOS in Rust and evaluate its performance across extensive ML environments, comparing it to state-of-the-art solutions from NVIDIA and prior research. For inference stacking, LithOS reduces tail latencies by 13x compared to MPS; compared to the best SotA, it reduces tail latencies by 3x while improving aggregate throughput by 1.6x. In hybrid inference-training stacking, LithOS reduces tail latencies by 4.7x compared to MPS; compared to the best SotA, it reduces tail latencies 1.18x while improving aggregate throughput by 1.35x. Finally, for a modest performance hit under 4%, LithOS's right-sizing provides a quarter of GPU capacity savings on average, while for a 7% hit, its power management yields a quarter of a GPU's energy savings. Overall, LithOS increases GPU efficiency, establishing a foundation for future OS research on GPUs.
We present our system submission for SemEval 2025 Task 5, which focuses on cross-lingual subject classification in the English and German academic domains. Our approach leverages bilingual data during training, employing negative sampling and a margin-based retrieval objective. We demonstrate that a dimension-as-token self-attention mechanism designed with significantly reduced internal dimensions can effectively encode sentence embeddings for subject retrieval. In quantitative evaluation, our system achieved an average recall rate of 32.24% in the general quantitative setting (all subjects), 43.16% and 31.53% of the general qualitative evaluation methods with minimal GPU usage, highlighting their competitive performance. Our results demonstrate that our approach is effective in capturing relevant subject information under resource constraints, although there is still room for improvement.
Heterogeneous deep learning systems (DLS) such as GPUs and ASICs have been widely deployed in industrial data centers, which requires to develop multiple low-level tensor programs for different platforms. An attractive solution to relieve the programming burden is to transcompile the legacy code of one platform to others. However, current transcompilation techniques struggle with either tremendous manual efforts or functional incorrectness, rendering "Write Once, Run Anywhere" of tensor programs an open question. We propose a novel transcompiler, i.e., QiMeng-Xpiler, for automatically translating tensor programs across DLS via both large language models (LLMs) and symbolic program synthesis, i.e., neural-symbolic synthesis. The key insight is leveraging the powerful code generation ability of LLM to make costly search-based symbolic synthesis computationally tractable. Concretely, we propose multiple LLM-assisted compilation passes via pre-defined meta-prompts for program transformation. During each program transformation, efficient symbolic program synthesis is employed to repair incorrect code snippets with a limited scale. To attain high performance, we propose a hierarchical auto-tuning approach to systematically explore both the parameters and sequences of transformation passes. Experiments on 4 DLS with distinct programming interfaces, i.e., Intel DL Boost with VNNI, NVIDIA GPU with CUDA, AMD MI with HIP, and Cambricon MLU with BANG, demonstrate that QiMeng-Xpiler correctly translates different tensor programs at the accuracy of 95% on average, and the performance of translated programs achieves up to 2.0x over vendor-provided manually-optimized libraries. As a result, the programming productivity of DLS is improved by up to 96.0x via transcompiling legacy tensor programs.
High-dimensional partial differential equations (PDEs) arise in diverse scientific and engineering applications but remain computationally intractable due to the curse of dimensionality. Traditional numerical methods struggle with the exponential growth in computational complexity, particularly on hypercubic domains, where the number of required collocation points increases rapidly with dimensionality. Here, we introduce Anant-Net, an efficient neural surrogate that overcomes this challenge, enabling the solution of PDEs in high dimensions. Unlike hyperspheres, where the internal volume diminishes as dimensionality increases, hypercubes retain or expand their volume (for unit or larger length), making high-dimensional computations significantly more demanding. Anant-Net efficiently incorporates high-dimensional boundary conditions and minimizes the PDE residual at high-dimensional collocation points. To enhance interpretability, we integrate Kolmogorov-Arnold networks into the Anant-Net architecture. We benchmark Anant-Net's performance on several linear and nonlinear high-dimensional equations, including the Poisson, Sine-Gordon, and Allen-Cahn equations, demonstrating high accuracy and robustness across randomly sampled test points from high-dimensional space. Importantly, Anant-Net achieves these results with remarkable efficiency, solving 300-dimensional problems on a single GPU within a few hours. We also compare Anant-Net's results for accuracy and runtime with other state-of-the-art methods. Our findings establish Anant-Net as an accurate, interpretable, and scalable framework for efficiently solving high-dimensional PDEs.
Anomaly detection in large datasets is essential in fields such as astronomy and computer vision; however, supervised methods typically require extensive anomaly labelling, which is often impractical. We present AnomalyMatch, an anomaly detection framework combining the semi-supervised FixMatch algorithm using EfficientNet classifiers with active learning. By treating anomaly detection as a semi-supervised binary classification problem, we efficiently utilise limited labelled and abundant unlabelled images. We allow iterative model refinement in a user interface for expert verification of high-confidence anomalies and correction of false positives. Built for astronomical data, AnomalyMatch generalises readily to other domains facing similar data challenges. Evaluations on the GalaxyMNIST astronomical dataset and the miniImageNet natural-image benchmark under severe class imbalance (1% anomalies for miniImageNet) display strong performance: starting from five to ten labelled anomalies and after three active learning cycles, we achieve an average AUROC of 0.95 (miniImageNet) and 0.86 (GalaxyMNIST), with respective AUPRC of 0.77 and 0.71. After active learning cycles, anomalies are ranked with 71% (miniImageNet) to 93% precision in the 1% of the highest-ranked images. AnomalyMatch is tailored for large-scale applications, efficiently processing predictions for 100 million images within three days on a single GPU. Integrated into ESAs Datalabs platform, AnomalyMatch facilitates targeted discovery of scientifically valuable anomalies in vast astronomical datasets. Our results underscore the exceptional utility and scalability of this approach for anomaly discovery, highlighting the value of specialised approaches for domains characterised by severe label scarcity.
Program synthesis is an umbrella term for generating programs and logical formulae from specifications. With the remarkable performance improvements that GPUs enable for deep learning, a natural question arose: can we also implement a search-based program synthesiser on GPUs to achieve similar performance improvements? In this article we discuss our insights on this question, based on recent works~. The goal is to build a synthesiser running on GPUs which takes as input positive and negative example traces and returns a logical formula accepting the positive and rejecting the negative traces. With GPU-friendly programming techniques -- using the semantics of formulae to minimise data movement and reduce data-dependent branching -- our synthesiser scales to significantly larger synthesis problems, and operates much faster than the previous CPU-based state-of-the-art. We believe the insights that make our approach GPU-friendly have wide potential for enhancing the performance of other formal methods (FM) workloads.
4D radar has attracted attention in autonomous driving due to its ability to enable robust 3D object detection even under adverse weather conditions. To practically deploy such technologies, it is essential to achieve real-time processing within low-power embedded environments. Addressing this, we present the first on-chip implementation of a 4D radar-based 3D object detection model on the Hailo-8L AI accelerator. Although conventional 3D convolutional neural network (CNN) architectures require 5D inputs, the Hailo-8L only supports 4D tensors, posing a significant challenge. To overcome this limitation, we introduce a tensor transformation method that reshapes 5D inputs into 4D formats during the compilation process, enabling direct deployment without altering the model structure. The proposed system achieves 46.47% AP_3D and 52.75% AP_BEV, maintaining comparable accuracy to GPU-based models while achieving an inference speed of 13.76 Hz. These results demonstrate the applicability of 4D radar-based perception technologies to autonomous driving systems.
Sionna is an open-source, GPU-accelerated library that, as of version 0.14, incorporates a ray tracer for simulating radio wave propagation. A unique feature of Sionna RT is differentiability, enabling the calculation of gradients for the channel impulse responses (CIRs), radio maps, and other related metrics with respect to system and environmental parameters, such as material properties, antenna patterns, and array geometries. The release of Sionna 1.0 provides a complete overhaul of the ray tracer, significantly improving its speed, memory efficiency, and extensibility. This document details the algorithms employed by Sionna RT to simulate radio wave propagation efficiently, while also addressing their current limitations. Given that the computation of CIRs and radio maps requires distinct algorithms, these are detailed in separate sections. For CIRs, Sionna RT integrates shooting and bouncing of rays (SBR) with the image method and uses a hashing-based mechanism to efficiently eliminate duplicate paths. Radio maps are computed using a purely SBR-based approach.
We present a method for Sparse view reconstruction with surface element splatting that runs within 3 minutes on a consumer grade GPU. While few methods address sparse radiance field learning from noisy or unposed sparse cameras, shape recovery remains relatively underexplored in this setting. Several radiance and shape learning test-time optimization methods address the sparse posed setting by learning data priors or using combinations of external monocular geometry priors. Differently, we propose an efficient and simple pipeline harnessing a single recent 3D foundation model. We leverage its various task heads, notably point maps and camera initializations to instantiate a bundle adjusting 2D Gaussian Splatting (2DGS) model, and image correspondences to guide camera optimization midst 2DGS training. Key to our contribution is a novel formulation of splatted color variance along rays, which can be computed efficiently. Reducing this moment in training leads to more accurate shape reconstructions. We demonstrate state-of-the-art performances in the sparse uncalibrated setting in reconstruction and novel view benchmarks based on established multi-view datasets.
Accurate mineral identification on the Martian surface is critical for understanding the planet's geological history. This paper presents a UNet-based autoencoder model for efficient spectral preprocessing of CRISM MTRDR hyperspectral data, addressing the limitations of traditional methods that are computationally intensive and time-consuming. The proposed model automates key preprocessing steps, such as smoothing and continuum removal, while preserving essential mineral absorption features. Trained on augmented spectra from the MICA spectral library, the model introduces realistic variability to simulate MTRDR data conditions. By integrating this framework, preprocessing time for an 800x800 MTRDR scene is reduced from 1.5 hours to just 5 minutes on an NVIDIA T1600 GPU. The preprocessed spectra are subsequently classified using MICAnet, a deep learning model for Martian mineral identification. Evaluation on labeled CRISM TRDR data demonstrates that the proposed approach achieves competitive accuracy while significantly enhancing preprocessing efficiency. This work highlights the potential of the UNet-based preprocessing framework to improve the speed and reliability of mineral mapping on Mars.
This paper lays the foundation for Genie, a testing framework that captures the impact of real hardware network behavior on ML workload performance, without requiring expensive GPUs. Genie uses CPU-initiated traffic over a hardware testbed to emulate GPU to GPU communication, and adapts the ASTRA-sim simulator to model interaction between the network and the ML workload.
Mixture-of-Experts (MoE) models have become a widely adopted solution to continue scaling model sizes without a corresponding linear increase in compute. During MoE model training, each input token is dynamically routed to a subset of experts -- sparsely-activated feed-forward networks -- within each transformer layer. The distribution of tokens assigned to each expert varies widely and rapidly over the course of training. To handle the wide load imbalance across experts, current systems are forced to either drop tokens assigned to popular experts, degrading convergence, or frequently rebalance resources allocated to each expert based on popularity, incurring high state migration overheads. To break this performance-accuracy tradeoff, we introduce SwiftMoE, an adaptive MoE training system. The key insight of SwiftMoE is to decouple the placement of expert parameters from their large optimizer state. SwiftMoE statically partitions the optimizer of each expert across all training nodes. Meanwhile, SwiftMoE dynamically adjusts the placement of expert parameters by repurposing existing weight updates, avoiding migration overheads. In doing so, SwiftMoE right-sizes the GPU resources allocated to each expert, on a per-iteration basis, with minimal overheads. Compared to state-of-the-art MoE training systems, DeepSpeed and FlexMoE, SwiftMoE is able to achieve a 30.5% and 25.9% faster time-to-convergence, respectively.
Existing large language model (LLM) serving systems fall into two categories: 1) a unified system where prefill phase and decode phase are co-located on the same GPU, sharing the unified computational resource and storage, and 2) a disaggregated system where the two phases are disaggregated to different GPUs. The design of the disaggregated system addresses the latency interference and sophisticated scheduling issues in the unified system but leads to storage challenges including 1) replicated weights for both phases that prevent flexible deployment, 2) KV cache transfer overhead between the two phases, 3) storage imbalance that causes substantial wasted space of the GPU capacity, and 4) suboptimal resource adjustment arising from the difficulties in migrating KV cache. Such storage inefficiency delivers poor serving performance under high request rates. In this paper, we identify that the advantage of the disaggregated system lies in the disaggregated computation, i.e., partitioning the computational resource to enable the asynchronous computation of two phases. Thus, we propose a novel LLM serving system, semi-PD, characterized by disaggregated computation and unified storage. In semi-PD, we introduce a computation resource controller to achieve disaggregated computation at the streaming multi-processor (SM) level, and a unified memory manager to manage the asynchronous memory access from both phases. semi-PD has a low-overhead resource adjustment mechanism between the two phases, and a service-level objective (SLO) aware dynamic partitioning algorithm to optimize the SLO attainment. Compared to state-of-the-art systems, semi-PD maintains lower latency at higher request rates, reducing the average end-to-end latency per request by 1.27-2.58x on DeepSeek series models, and serves 1.55-1.72x more requests adhering to latency constraints on Llama series models.
Large Language Models (LLMs) for Generative AI have achieved remarkable progress, evolving into sophisticated and versatile tools widely adopted across various domains and applications. However, the substantial memory overhead caused by their vast number of parameters, combined with the high computational demands of the attention mechanism, poses significant challenges in achieving low latency and high throughput for LLM inference services. Recent advancements, driven by groundbreaking research, have significantly accelerated progress in this field. This paper provides a comprehensive survey of these methods, covering fundamental instance-level approaches, in-depth cluster-level strategies, emerging scenario directions, and other miscellaneous but important areas. At the instance level, we review model placement, request scheduling, decoding length prediction, storage management, and the disaggregation paradigm. At the cluster level, we explore GPU cluster deployment, multi-instance load balancing, and cloud service solutions. For emerging scenarios, we organize the discussion around specific tasks, modules, and auxiliary methods. To ensure a holistic overview, we also highlight several niche yet critical areas. Finally, we outline potential research directions to further advance the field of LLM inference serving.
Generative models have achieved remarkable success across various applications, driving the demand for multi-GPU computing. Inter-GPU communication becomes a bottleneck in multi-GPU computing systems, particularly on consumer-grade GPUs. By exploiting concurrent hardware execution, overlapping computation and communication latency is an effective technique for mitigating the communication overhead. We identify that an efficient and adaptable overlapping design should satisfy (1) tile-wise overlapping to maximize the overlapping opportunity, (2) interference-free computation to maintain the original computational performance, and (3) communication agnosticism to reduce the development burden against varying communication primitives. Nevertheless, current designs fail to simultaneously optimize for all of those features. To address the issue, we propose FlashOverlap, a lightweight design characterized by tile-wise overlapping, interference-free computation, and communication agnosticism. FlashOverlap utilizes a novel signaling mechanism to identify tile-wise data dependency without interrupting the computation process, and reorders data to contiguous addresses, enabling communication by simply calling NCCL APIs. Experiments show that such a lightweight design achieves up to 1.65x speedup, outperforming existing works in most cases.
We evaluate the current state of collective communication on GPU-based supercomputers for large language model (LLM) training at scale. Existing libraries such as RCCL and Cray-MPICH exhibit critical limitations on systems such as Frontier -- Cray-MPICH underutilizes network and compute resources, while RCCL suffers from severe scalability issues. To address these challenges, we introduce PCCL, a communication library with highly optimized implementations of all-gather and reduce-scatter operations tailored for distributed deep learning workloads. PCCL is designed to maximally utilize all available network and compute resources and to scale efficiently to thousands of GPUs. It achieves substantial performance improvements, delivering 6-33x speedups over RCCL and 28-70x over Cray-MPICH for all-gather on 2048 GCDs of Frontier. These gains translate directly to end-to-end performance: in large-scale GPT-3-style training, PCCL provides up to 60% and 40% speedups over RCCL for 7B and 13B parameter models, respectively.