With the ever-growing size of real-world graphs, numerous techniques to overcome resource limitations when training Graph Neural Networks (GNNs) have been developed. One such approach, GNNAutoScale (GAS), uses graph partitioning to enable training under constrained GPU memory. GAS also stores historical embedding vectors, which are retrieved from one-hop neighbors in other partitions, ensuring critical information is captured across partition boundaries. The historical embeddings which come from the previous training iteration are stale compared to the GAS estimated embeddings, resulting in approximation errors of the training algorithm. Furthermore, these errors accumulate over multiple layers, leading to suboptimal node embeddings. To address this shortcoming, we propose two enhancements: first, WaveGAS, inspired by waveform relaxation, performs multiple forward passes within GAS before the backward pass, refining the approximation of historical embeddings and gradients to improve accuracy; second, a gradient-tracking method that stores and utilizes more accurate historical gradients during training. Empirical results show that WaveGAS enhances GAS and achieves better accuracy, even outperforming methods that train on full graphs, thanks to its robust estimation of node embeddings.
Message passing-based graph neural networks (GNNs) have achieved great success in many real-world applications. For a sampled mini-batch of target nodes, the message passing process is divided into two parts: message passing between nodes within the batch (MP-IB) and message passing from nodes outside the batch to those within it (MP-OB). However, MP-OB recursively relies on higher-order out-of-batch neighbors, leading to an exponentially growing computational cost with respect to the number of layers. Due to the neighbor explosion, the whole message passing stores most nodes and edges on the GPU such that many GNNs are infeasible to large-scale graphs. To address this challenge, we propose an accurate and fast mini-batch approach for large graph transductive learning, namely topological compensation (TOP), which obtains the outputs of the whole message passing solely through MP-IB, without the costly MP-OB. The major pillar of TOP is a novel concept of message invariance, which defines message-invariant transformations to convert costly MP-OB into fast MP-IB. This ensures that the modified MP-IB has the same output as the whole message passing. Experiments demonstrate that TOP is significantly faster than existing mini-batch methods by order of magnitude on vast graphs (millions of nodes and billions of edges) with limited accuracy degradation.
We implement a specialized version of our SpeckleNN model for real-time speckle pattern classification in X-ray Single-Particle Imaging (SPI) using the SLAC Neural Network Library (SNL) on an FPGA. This hardware is optimized for inference near detectors in high-throughput X-ray free-electron laser (XFEL) facilities like the Linac Coherent Light Source (LCLS). To fit FPGA constraints, we optimized SpeckleNN, reducing parameters from 5.6M to 64.6K (98.8% reduction) with 90% accuracy. We also compressed the latent space from 128 to 50 dimensions. Deployed on a KCU1500 FPGA, the model used 71% of DSPs, 75% of LUTs, and 48% of FFs, with an average power consumption of 9.4W. The FPGA achieved 45.015us inference latency at 200 MHz. On an NVIDIA A100 GPU, the same inference consumed ~73W and had a 400us latency. Our FPGA version achieved an 8.9x speedup and 7.8x power reduction over the GPU. Key advancements include model specialization and dynamic weight loading through SNL, eliminating time-consuming FPGA re-synthesis for fast, continuous deployment of (re)trained models. These innovations enable real-time adaptive classification and efficient speckle pattern vetoing, making SpeckleNN ideal for XFEL facilities. This implementation accelerates SPI experiments and enhances adaptability to evolving conditions.
The signature kernel is a recent state-of-the-art tool for analyzing high-dimensional sequential data, valued for its theoretical guarantees and strong empirical performance. In this paper, we present a novel method for efficiently computing the signature kernel of long, high-dimensional time series via dynamically truncated recursive local power series expansions. Building on the characterization of the signature kernel as the solution of a Goursat PDE, our approach employs tilewise Neumann-series expansions to derive rapidly converging power series approximations of the signature kernel that are locally defined on subdomains and propagated iteratively across the entire domain of the Goursat solution by exploiting the geometry of the time series. Algorithmically, this involves solving a system of interdependent local Goursat PDEs by recursively propagating boundary conditions along a directed graph via topological ordering, with dynamic truncation adaptively terminating each local power series expansion when coefficients fall below machine precision, striking an effective balance between computational cost and accuracy. This method achieves substantial performance improvements over state-of-the-art approaches for computing the signature kernel, providing (a) adjustable and superior accuracy, even for time series with very high roughness; (b) drastically reduced memory requirements; and (c) scalability to efficiently handle very long time series (e.g., with up to half a million points or more) on a single GPU. These advantages make our method particularly well-suited for rough-path-assisted machine learning, financial modeling, and signal processing applications that involve very long and highly volatile data.
Mixture-of-experts (MoE) has been extensively employed to scale large language models to trillion-plus parameters while maintaining a fixed computational cost. The development of large MoE models in the distributed scenario encounters the problem of large communication overhead. The inter-device communication of a MoE layer can occupy 47% time of the entire model execution with popular models and frameworks. Therefore, existing methods suggest the communication in a MoE layer to be pipelined with the computation for overlapping. However, these coarse grained overlapping schemes introduce a notable impairment of computational efficiency and the latency concealing is sub-optimal. To this end, we present COMET, an optimized MoE system with fine-grained communication-computation overlapping. Leveraging data dependency analysis and task rescheduling, COMET achieves precise fine-grained overlapping of communication and computation. Through adaptive workload assignment, COMET effectively eliminates fine-grained communication bottlenecks and enhances its adaptability across various scenarios. Our evaluation shows that COMET accelerates the execution of a single MoE layer by $1.96\times$ and for end-to-end execution, COMET delivers a $1.71\times$ speedup on average. COMET has been adopted in the production environment of clusters with ten-thousand-scale of GPUs, achieving savings of millions of GPU hours.
We present Mechanistic PDE Networks -- a model for discovery of governing partial differential equations from data. Mechanistic PDE Networks represent spatiotemporal data as space-time dependent linear partial differential equations in neural network hidden representations. The represented PDEs are then solved and decoded for specific tasks. The learned PDE representations naturally express the spatiotemporal dynamics in data in neural network hidden space, enabling increased power for dynamical modeling. Solving the PDE representations in a compute and memory-efficient way, however, is a significant challenge. We develop a native, GPU-capable, parallel, sparse, and differentiable multigrid solver specialized for linear partial differential equations that acts as a module in Mechanistic PDE Networks. Leveraging the PDE solver, we propose a discovery architecture that can discover nonlinear PDEs in complex settings while also being robust to noise. We validate PDE discovery on a number of PDEs, including reaction-diffusion and Navier-Stokes equations.
Placement is a critical task with high computation complexity in VLSI physical design. Modern analytical placers formulate the placement objective as a nonlinear optimization task, which suffers a long iteration time. To accelerate and enhance the placement process, recent studies have turned to deep learning-based approaches, particularly leveraging graph convolution networks (GCNs). However, learning-based placers require time- and data-consuming model training due to the complexity of circuit placement that involves large-scale cells and design-specific graph statistics. This paper proposes GiFt, a parameter-free technique for accelerating placement, rooted in graph signal processing. GiFt excels at capturing multi-resolution smooth signals of circuit graphs to generate optimized placement solutions without the need for time-consuming model training, and meanwhile significantly reduces the number of iterations required by analytical placers. Experimental results show that GiFt significantly improving placement efficiency, while achieving competitive or superior performance compared to state-of-the-art placers. In particular, compared to DREAMPlace, the recently proposed GPU-accelerated analytical placer, GF-Placer improves total runtime over 45%.
Accurate click-through rate (CTR) prediction is vital for online advertising and recommendation systems. Recent deep learning advancements have improved the ability to capture feature interactions and understand user interests. However, optimizing the embedding layer often remains overlooked. Embedding tables, which represent categorical and sequential features, can become excessively large, surpassing GPU memory limits and necessitating storage in CPU memory. This results in high memory consumption and increased latency due to frequent GPU-CPU data transfers. To tackle these challenges, we introduce a Model-agnostic Embedding Compression (MEC) framework that compresses embedding tables by quantizing pre-trained embeddings, without sacrificing recommendation quality. Our approach consists of two stages: first, we apply popularity-weighted regularization to balance code distribution between high- and low-frequency features. Then, we integrate a contrastive learning mechanism to ensure a uniform distribution of quantized codes, enhancing the distinctiveness of embeddings. Experiments on three datasets reveal that our method reduces memory usage by over 50x while maintaining or improving recommendation performance compared to existing models. The implementation code is accessible in our project repository https://github.com/USTC-StarTeam/MEC.
In recent years, more and more attention has been paid to the use of artificial neural networks (ANN) for diagnostics of gas pumping units (GPU). Usually, ANN training is carried out on models of GPU workflows, and generated sets of diagnostic data are used to simulate defect conditions. At the same time, the results obtained do not allow assessing the real state of the GPU. It is proposed to use the values of the characteristics of the acoustic and vibration processes of the GPU as the input data of the ANN. A descriptive statistical analysis of real vibration and acoustic processes generated by the operation of the GPU type GTK-25-i (Nuovo Pignone, Italy) has been carried out. The formation of packets of diagnostic signs arriving at the input of the ANN has been carried out. The diagnostic features are the five maximum amplitude components of the acoustic and vibration signals, as well as the value of the standard deviation for each sample. Diagnostic signs are calculated directly in the input pipeline of ANN data in real time for three technical states of the GPU. Using the frameworks TensorFlow, Keras, NumPy, pandas, in the Python 3 programming language, an architecture was developed for a deep fully connected feedforward ANN, training on the error backpropagation algorithm. The results of training and testing of the developed ANN are presented. During testing, it was found that the signal classification precision for the "nominal" state of all 1475 signal samples is 1.0000, for the "current" state, precision equils 0.9853, and for the "defective" state, precision is 0.9091. The use of the developed ANN makes it possible to classify the technical states of the GPU with an accuracy sufficient for practical use, which will prevent the occurrence of GPU failures. ANN can be used to diagnose GPU of any type and power.
Training on edge devices poses several challenges as these devices are generally resource-constrained, especially in terms of power. State-of-the-art techniques at the device level reduce the GPU frequency to enforce power constraints, leading to a significant increase in training time. To accelerate training, we propose to jointly adjust the system and application parameters (in our case, the GPU frequency and the batch size of the training task) while adhering to the power constraints on devices. We introduce a novel cross-layer methodology that combines predictions of batch size efficiency and device profiling to achieve the desired optimization. Our evaluation on real hardware shows that our method outperforms the current baselines that depend on state of the art techniques, reducing the training time by $2.4\times$ with results very close to optimal. Our measurements also indicate a substantial reduction in the overall energy used for the training process. These gains are achieved without reduction in the performance of the trained model.
Nvidia Jetson boards are powerful systems for executing artificial intelligence workloads in edge and mobile environments due to their effective GPU hardware and widely supported software stack. In addition to these benefits, Nvidia Jetson boards provide large configurability by giving the user the choice to modify many hardware parameters. This large space of configurability creates the need of searching the optimal configurations based on the user's requirements. In this work, we propose JExplore, a multi-board software and hardware design space exploration tool. JExplore can be integrated with any search tool, hence creating a common benchmarking ground for the search algorithms. Moreover, it accelerates the exploration of user application and Nvidia Jetson configurations for researchers and engineers by encapsulating host-client communication, configuration management, and metric measurement.
Knowledge graph (KG) learning offers a powerful framework for generating new knowledge and making inferences. Training KG embedding can take a significantly long time, especially for larger datasets. Our analysis shows that the gradient computation of embedding is one of the dominant functions in the translation-based KG embedding training loop. We address this issue by replacing the core embedding computation with SpMM (Sparse-Dense Matrix Multiplication) kernels. This allows us to unify multiple scatter (and gather) operations as a single operation, reducing training time and memory usage. We create a general framework for training KG models using sparse kernels and implement four models, namely TransE, TransR, TransH, and TorusE. Our sparse implementations exhibit up to 5.3x speedup on the CPU and up to 4.2x speedup on the GPU with a significantly low GPU memory footprint. The speedups are consistent across large and small datasets for a given model. Our proposed sparse approach can be extended to accelerate other translation-based (such as TransC, TransM, etc.) and non-translational (such as DistMult, ComplEx, RotatE, etc.) models as well. An implementation of the SpTransX framework is publicly available as a Python package in https://github.com/HipGraph/SpTransX.
Recent studies show that the visual place recognition (VPR) method using pre-trained visual foundation models can achieve promising performance. In our previous work, we propose a novel method to realize seamless adaptation of foundation models to VPR (SelaVPR). This method can produce both global and local features that focus on discriminative landmarks to recognize places for two-stage VPR by a parameter-efficient adaptation approach. Although SelaVPR has achieved competitive results, we argue that the previous adaptation is inefficient in training time and GPU memory usage, and the re-ranking paradigm is also costly in retrieval latency and storage usage. In pursuit of higher efficiency and better performance, we propose an extension of the SelaVPR, called SelaVPR++. Concretely, we first design a parameter-, time-, and memory-efficient adaptation method that uses lightweight multi-scale convolution (MultiConv) adapters to refine intermediate features from the frozen foundation backbone. This adaptation method does not back-propagate gradients through the backbone during training, and the MultiConv adapter facilitates feature interactions along the spatial axes and introduces proper local priors, thus achieving higher efficiency and better performance. Moreover, we propose an innovative re-ranking paradigm for more efficient VPR. Instead of relying on local features for re-ranking, which incurs huge overhead in latency and storage, we employ compact binary features for initial retrieval and robust floating-point (global) features for re-ranking. To obtain such binary features, we propose a similarity-constrained deep hashing method, which can be easily integrated into the VPR pipeline. Finally, we improve our training strategy and unify the training protocol of several common training datasets to merge them for better training of VPR models. Extensive experiments show that ......
Embedding tree-like data, from hierarchies to ontologies and taxonomies, forms a well-studied problem for representing knowledge across many domains. Hyperbolic geometry provides a natural solution for embedding trees, with vastly superior performance over Euclidean embeddings. Recent literature has shown that hyperbolic tree embeddings can even be placed on top of neural networks for hierarchical knowledge integration in deep learning settings. For all applications, a faithful embedding of trees is needed, with combinatorial constructions emerging as the most effective direction. This paper identifies and solves two key limitations of existing works. First, the combinatorial construction hinges on finding highly separated points on a hypersphere, a notoriously difficult problem. Current approaches achieve poor separation, degrading the quality of the corresponding hyperbolic embedding. We propose highly separated Delaunay tree embeddings (HS-DTE), which integrates angular separation in a generalized formulation of Delaunay embeddings, leading to lower embedding distortion. Second, low-distortion requires additional precision. The current approach for increasing precision is to use multiple precision arithmetic, which renders the embeddings useless on GPUs in deep learning settings. We reformulate the combinatorial construction using floating point expansion arithmetic, leading to superior embedding quality while retaining utility on accelerated hardware.
Retrieval Augmented Generation (RAG) systems remain vulnerable to hallucinated answers despite incorporating external knowledge sources. We present LettuceDetect a framework that addresses two critical limitations in existing hallucination detection methods: (1) the context window constraints of traditional encoder-based methods, and (2) the computational inefficiency of LLM based approaches. Building on ModernBERT's extended context capabilities (up to 8k tokens) and trained on the RAGTruth benchmark dataset, our approach outperforms all previous encoder-based models and most prompt-based models, while being approximately 30 times smaller than the best models. LettuceDetect is a token-classification model that processes context-question-answer triples, allowing for the identification of unsupported claims at the token level. Evaluations on the RAGTruth corpus demonstrate an F1 score of 79.22% for example-level detection, which is a 14.8% improvement over Luna, the previous state-of-the-art encoder-based architecture. Additionally, the system can process 30 to 60 examples per second on a single GPU, making it more practical for real-world RAG applications.
Proximal Policy Optimization (PPO)-based Reinforcement Learning from Human Feedback (RLHF) is essential for aligning large language models (LLMs) with human preferences. It requires joint training of an actor and critic with a pretrained, fixed reward model for guidance. This approach increases computational complexity and instability due to actor-critic interdependence. Additionally, PPO lacks access to true environment rewards in LLM tasks, limiting its adaptability. Under such conditions, pretraining a value model or a reward model becomes equivalent, as both provide fixed supervisory signals without new ground-truth feedback. To address these issues, we propose \textbf{Decoupled Value Policy Optimization (DVPO)}, a lean framework that replaces traditional reward modeling with a pretrained \emph{global value model (GVM)}. The GVM is conditioned on policy trajectories and predicts token-level return-to-go estimates. By decoupling value model from policy training (via frozen GVM-driven RL objectives), DVPO eliminates actor-critic interdependence, reducing GPU memory usage by 40\% and training time by 35\% compared to conventional RLHF. Experiments across benchmarks show DVPO outperforms efficient RLHF methods (e.g., DPO) while matching state-of-the-art PPO in performance.
The increasing context window size in large language models (LLMs) has improved their ability to handle complex, long-text tasks. However, as the conversation rounds continue, it is required to store a large amount of KV cache in GPU memory, which significantly affects the efficiency and even availability of the model serving systems. This paper analyzes dialogue data from real users and discovers that the LLM inference manifests a watershed layer, after which the distribution of round-level attention shows notable similarity. We propose Round Attention, a novel round-level attention mechanism that only recalls and computes the KV cache of the most relevant rounds. The experiments show that our method saves 55\% memory usage without compromising model performance.
This paper introduces a new Transformer, called MS$^2$Dformer, that can be used as a generalized backbone for multi-modal sequence spammer detection. Spammer detection is a complex multi-modal task, thus the challenges of applying Transformer are two-fold. Firstly, complex multi-modal noisy information about users can interfere with feature mining. Secondly, the long sequence of users' historical behaviors also puts a huge GPU memory pressure on the attention computation. To solve these problems, we first design a user behavior Tokenization algorithm based on the multi-modal variational autoencoder (MVAE). Subsequently, a hierarchical split-window multi-head attention (SW/W-MHA) mechanism is proposed. The split-window strategy transforms the ultra-long sequences hierarchically into a combination of intra-window short-term and inter-window overall attention. Pre-trained on the public datasets, MS$^2$Dformer's performance far exceeds the previous state of the art. The experiments demonstrate MS$^2$Dformer's ability to act as a backbone.
The Qwen 2.5 3B base model was fine-tuned to generate contextually rich and engaging movie dialogue, leveraging the Cornell Movie-Dialog Corpus, a curated dataset of movie conversations. Due to the limitations in GPU computing and VRAM, the training process began with the 0.5B model progressively scaling up to the 1.5B and 3B versions as efficiency improvements were implemented. The Qwen 2.5 series, developed by Alibaba Group, stands at the forefront of small open-source pre-trained models, particularly excelling in creative tasks compared to alternatives like Meta's Llama 3.2 and Google's Gemma. Results demonstrate the ability of small models to produce high-quality, realistic dialogue, offering a promising approach for real-time, context-sensitive conversation generation.
KV cache techniques in Transformer models aim to reduce redundant computations at the expense of substantially increased memory usage, making KV cache compression an important and popular research topic. Recently, state-of-the-art KV cache compression methods implement imbalanced, per-head allocation algorithms that dynamically adjust the KV cache budget for each attention head, achieving excellent performance in single-GPU scenarios. However, we observe that such imbalanced compression leads to significant load imbalance when deploying multi-GPU inference, as some GPUs become overburdened while others remain underutilized. In this paper, we propose FairKV, a method designed to ensure fair memory usage among attention heads in systems employing imbalanced KV cache compression. The core technique of FairKV is Fair-Copying, which replicates a small subset of memory-intensive attention heads across GPUs using data parallelism to mitigate load imbalance. Our experiments on popular models, including LLaMA 70b and Mistral 24b model, demonstrate that FairKV increases throughput by 1.66x compared to standard tensor parallelism inference. Our code will be released as open source upon acceptance.