Communication overhead in federated learning (FL) poses a significant challenge for network anomaly detection systems, where diverse client configurations and network conditions impact efficiency and detection accuracy. Existing approaches attempt optimization individually but struggle to balance reduced overhead with performance. This paper presents an adaptive FL framework combining batch size optimization, client selection, and asynchronous updates for efficient anomaly detection. Using UNSW-NB15 for general network traffic and ROAD for automotive networks, our framework reduces communication overhead by 97.6% (700.0s to 16.8s) while maintaining comparable accuracy (95.10% vs. 95.12%). The Mann-Whitney U test confirms significant improvements (p < 0.05). Profiling analysis reveals efficiency gains via reduced GPU operations and memory transfers, ensuring robust detection across varying client conditions.
In the era of LLMs, dense operations such as GEMM and MHA are critical components. These operations are well-suited for parallel execution using a tilebased approach. While traditional GPU programming often relies on low level interfaces like CUDA or SYCL, Triton has emerged as a DSL that offers a more user-friendly and portable alternative by programming at a higher level. The current Triton starts at the workgroup (aka threadblock) level, and directly lowers to per-thread level. And then attempt to coalesce and amend through a series of passes, promoting information from low-level representation. We believe this is pre-mature lowering based on the below observations. 1. GPU has a hierarchical structure both physically and logically. Modern GPUs often feature SIMD units capable of directly operating on tiles on a warp or warpgroup basis, such as blocked load and blocked MMA. 2. Multi-level gradual lowering can make compiler decoupled and clean by separating considerations inter and intra a logical layer. 3. Kernel developers often need fine control to get good performance on the latest hardware. FlashAttention2 advocates explicit data partition between warps to make a performance boost. In this context, we propose ML-Triton which features multi-level compilation flow and programming interface. Our approach begins at the workgroup level and progressively lowers to the warp and intrinsic level, implementing a multilevel lowering align with the hierarchical nature of GPU. Additionally, we extend triton language to support user-set compiler hint and warp level programming, enabling researchers to get good out-of-the box performance without awaiting compiler updates. Experimental results demonstrate that our approach achieves performance above 95% of expert-written kernels on Intel GPU, as measured by the geometric mean.
In this paper, we consider a mixed-prompt scenario for a large language model (LLM) inference serving system that supports diverse applications with both short prompts and long prompts and heterogeneous SLOs for iteration time. To improve throughput when handling long prompts, previous research introduces a chunking method, but has not addressed heterogeneous SLOs. To address the limitation, we propose AccelGen, a high-throughput LLM inference serving system with heterogeneous SLO guarantees for diverse applications. AccelGen introduces four core components: (1) SLO-guaranteed dynamic chunking, which dynamically adjusts chunk sizes to maximize GPU compute utilization while meeting iteration-level SLOs; (2) Iteration-level SLO-based task prioritization, which prioritizes tight-SLO requests and batches requests with similar SLOs; (3) Multi-resource-aware batching, which selects queued requests to maximize the utilizations of both GPU compute resource and key-value cache (KVC). Trace-driven real experiments demonstrate that AccelGen achieves 1.42-11.21X higher throughput, 1.43-13.71X higher goodput, 37-90% higher SLO attainment, and 1.61-12.22X lower response latency compared to the state-of-the-art approaches. It achieves performance near the Oracle, which optimally maximizes goodput.
Self-supervised 3D occupancy prediction offers a promising solution for understanding complex driving scenes without requiring costly 3D annotations. However, training dense voxel decoders to capture fine-grained geometry and semantics can demand hundreds of GPU hours, and such models often fail to adapt to varying voxel resolutions or new classes without extensive retraining. To overcome these limitations, we propose a practical and flexible test-time occupancy prediction framework termed TT-GaussOcc. Our approach incrementally optimizes time-aware 3D Gaussians instantiated from raw sensor streams at runtime, enabling voxelization at arbitrary user-specified resolution. Specifically, TT-GaussOcc operates in a "lift-move-voxel" symphony: we first "lift" surrounding-view semantics obtained from 2D vision foundation models (VLMs) to instantiate Gaussians at non-empty 3D space; Next, we "move" dynamic Gaussians from previous frames along estimated Gaussian scene flow to complete appearance and eliminate trailing artifacts of fast-moving objects, while accumulating static Gaussians to enforce temporal consistency; Finally, we mitigate inherent noises in semantic predictions and scene flow vectors by periodically smoothing neighboring Gaussians during optimization, using proposed trilateral RBF kernels that jointly consider color, semantic, and spatial affinities. The historical static and current dynamic Gaussians are then combined and voxelized to generate occupancy prediction. Extensive experiments on Occ3D and nuCraft with varying voxel resolutions demonstrate that TT-GaussOcc surpasses self-supervised baselines by 46% on mIoU without any offline training, and supports finer voxel resolutions at 2.6 FPS inference speed.
Linear RNNs with gating recently demonstrated competitive performance compared to Transformers in language modeling. Although their linear compute scaling in sequence length offers theoretical runtime advantages over Transformers, realizing these benefits in practice requires optimized custom kernels, as Transformers rely on the highly efficient Flash Attention kernels (Dao, 2024). Leveraging the chunkwise-parallel formulation of linear RNNs, Flash Linear Attention (FLA) (Yang & Zhang, 2024) shows that linear RNN kernels are faster than Flash Attention, by parallelizing over chunks of the input sequence. However, since the chunk size of FLA is limited, many intermediate states must be materialized in GPU memory. This leads to low arithmetic intensity and causes high memory consumption and IO cost, especially for long-context pre-training. In this work, we present Tiled Flash Linear Attention (TFLA), a novel kernel algorithm for linear RNNs, that enables arbitrary large chunk sizes and high arithmetic intensity by introducing an additional level of sequence parallelization within each chunk. First, we apply TFLA to the xLSTM with matrix memory, the mLSTM (Beck et al., 2024). Second, we propose an mLSTM variant with sigmoid input gate and reduced computation for even faster kernel runtimes at equal language modeling performance. In our speed benchmarks, we show that our new mLSTM kernels based on TFLA outperform highly optimized Flash Attention, Linear Attention and Mamba kernels, setting a new state of the art for efficient long-context sequence modeling primitives.
Leading large language models have demonstrated impressive capabilities in reasoning-intensive tasks, such as standardized educational testing. However, they often require extensive training in low-resource settings with inaccessible infrastructure. Small or compact models, though more efficient, frequently lack sufficient support for underrepresented languages, leaving a performance gap in critical domains. This work explores the potential of parameter-efficient fine-tuning of compact open-weight language models to handle reasoning-intensive tasks in the underrepresented Ukrainian language, building on the findings of the ZNO-Eval benchmark. Parameter-efficient fine-tuning of LLaMA 3.1 (8 billion parameters), LLaMA 3.2 (3 billion parameters), and Gemma 2 (9 billion parameters) models on chain-of-thought solutions resulted in a modest test score improvement of up to 17.4% on complex matching tasks and 1.6% overall compared to tuning on answer letters alone, offering enhanced interpretability and robustness. In addition, the proposed tuning method with joint task topic and step-by-step solution generation outperforms standard chain-of-thought tuning in matching tasks and provides a 5.4% gain over the best LLaMA 3.2 model due to guiding the model to recall and apply domain-relevant information. Contrasting obtained results with zero-shot evaluations of leading open-weight and proprietary models such as Qwen, DeepSeek R1, OpenAI o1 and o3, Gemini, and Claude, highlight that fine-tuning LLaMA and Gemma models with 2,032 step-by-step solutions and 20 to 50 million trainable parameters on a single A100 GPU lets them outperform GPT-4o mini, Mistral Large, and larger open-weight models. This research also evaluates how merging the quantized adapter with the base model influences the generation quality. Source code and tuned models are available at https://github.com/NLPForUA/ZNO.
We present a latent diffusion model for fast feed-forward 3D scene generation. Given one or more images, our model Bolt3D directly samples a 3D scene representation in less than seven seconds on a single GPU. We achieve this by leveraging powerful and scalable existing 2D diffusion network architectures to produce consistent high-fidelity 3D scene representations. To train this model, we create a large-scale multiview-consistent dataset of 3D geometry and appearance by applying state-of-the-art dense 3D reconstruction techniques to existing multiview image datasets. Compared to prior multiview generative models that require per-scene optimization for 3D reconstruction, Bolt3D reduces the inference cost by a factor of up to 300 times.
Top-leading solutions for Video Scene Graph Generation (VSGG) typically adopt an offline pipeline. Though demonstrating promising performance, they remain unable to handle real-time video streams and consume large GPU memory. Moreover, these approaches fall short in temporal reasoning, merely aggregating frame-level predictions over a temporal context. In response, we introduce DIFFVSGG, an online VSGG solution that frames this task as an iterative scene graph update problem. Drawing inspiration from Latent Diffusion Models (LDMs) which generate images via denoising a latent feature embedding, we unify the decoding of object classification, bounding box regression, and graph generation three tasks using one shared feature embedding. Then, given an embedding containing unified features of object pairs, we conduct a step-wise Denoising on it within LDMs, so as to deliver a clean embedding which clearly indicates the relationships between objects. This embedding then serves as the input to task-specific heads for object classification, scene graph generation, etc. DIFFVSGG further facilitates continuous temporal reasoning, where predictions for subsequent frames leverage results of past frames as the conditional inputs of LDMs, to guide the reverse diffusion process for current frames. Extensive experiments on three setups of Action Genome demonstrate the superiority of DIFFVSGG.
Video object detection has made significant progress in recent years thanks to convolutional neural networks (CNNs) and vision transformers (ViTs). Typically, CNNs excel at capturing local features but struggle to model global representations. Conversely, ViTs are adept at capturing long-range global features but face challenges in representing local feature details. Off-the-shelf video object detection methods solely rely on CNNs or ViTs to conduct feature aggregation, which hampers their capability to simultaneously leverage global and local information, thereby resulting in limited detection performance. In this paper, we propose a Transformer-GraphFormer Blender Network (TGBFormer) for video object detection, with three key technical improvements to fully exploit the advantages of transformers and graph convolutional networks while compensating for their limitations. First, we develop a spatial-temporal transformer module to aggregate global contextual information, constituting global representations with long-range feature dependencies. Second, we introduce a spatial-temporal GraphFormer module that utilizes local spatial and temporal relationships to aggregate features, generating new local representations that are complementary to the transformer outputs. Third, we design a global-local feature blender module to adaptively couple transformer-based global representations and GraphFormer-based local representations. Extensive experiments demonstrate that our TGBFormer establishes new state-of-the-art results on the ImageNet VID dataset. Particularly, our TGBFormer achieves 86.5% mAP while running at around 41.0 FPS on a single Tesla A100 GPU.
Long-form video understanding is essential for various applications such as video retrieval, summarizing, and question answering. Yet, traditional approaches demand substantial computing power and are often bottlenecked by GPU memory. To tackle this issue, we present Long-Video Memory Network, Long-VMNet, a novel video understanding method that employs a fixed-size memory representation to store discriminative patches sampled from the input video. Long-VMNet achieves improved efficiency by leveraging a neural sampler that identifies discriminative tokens. Additionally, Long-VMNet only needs one scan through the video, greatly boosting efficiency. Our results on the Rest-ADL dataset demonstrate an 18x -- 75x improvement in inference times for long-form video retrieval and answering questions, with a competitive predictive performance.
Despite the substantial advancements demonstrated by learning-based neural models in the LiDAR Point Cloud Compression (LPCC) task, realizing real-time compression - an indispensable criterion for numerous industrial applications - remains a formidable challenge. This paper proposes RENO, the first real-time neural codec for 3D LiDAR point clouds, achieving superior performance with a lightweight model. RENO skips the octree construction and directly builds upon the multiscale sparse tensor representation. Instead of the multi-stage inferring, RENO devises sparse occupancy codes, which exploit cross-scale correlation and derive voxels' occupancy in a one-shot manner, greatly saving processing time. Experimental results demonstrate that the proposed RENO achieves real-time coding speed, 10 fps at 14-bit depth on a desktop platform (e.g., one RTX 3090 GPU) for both encoding and decoding processes, while providing 12.25% and 48.34% bit-rate savings compared to G-PCCv23 and Draco, respectively, at a similar quality. RENO model size is merely 1MB, making it attractive for practical applications. The source code is available at https://github.com/NJUVISION/RENO.
The field of Novel View Synthesis has been revolutionized by 3D Gaussian Splatting (3DGS), which enables high-quality scene reconstruction that can be rendered in real-time. 3DGS-based techniques typically suffer from high GPU memory and disk storage requirements which limits their practical application on consumer-grade devices. We propose Opti3DGS, a novel frequency-modulated coarse-to-fine optimization framework that aims to minimize the number of Gaussian primitives used to represent a scene, thus reducing memory and storage demands. Opti3DGS leverages image frequency modulation, initially enforcing a coarse scene representation and progressively refining it by modulating frequency details in the training images. On the baseline 3DGS, we demonstrate an average reduction of 62% in Gaussians, a 40% reduction in the training GPU memory requirements and a 20% reduction in optimization time without sacrificing the visual quality. Furthermore, we show that our method integrates seamlessly with many 3DGS-based techniques, consistently reducing the number of Gaussian primitives while maintaining, and often improving, visual quality. Additionally, Opti3DGS inherently produces a level-of-detail scene representation at no extra cost, a natural byproduct of the optimization pipeline. Results and code will be made publicly available.
Fine-tuning large pre-trained LLMs generally demands extensive GPU memory. Traditional first-order optimizers like SGD encounter substantial difficulties due to increased memory requirements from storing activations and gradients during both the forward and backward phases as the model size expands. Alternatively, zeroth-order (ZO) techniques can compute gradients using just forward operations, eliminating the need to store activations. Furthermore, by leveraging CPU capabilities, it's feasible to enhance both the memory and processing power available to a single GPU. We propose a novel framework, ZO2 (Zeroth-Order Offloading), for efficient zeroth-order fine-tuning of LLMs with only limited GPU memory. Our framework dynamically shifts model parameters between the CPU and GPU as required, optimizing computation flow and maximizing GPU usage by minimizing downtime. This integration of parameter adjustments with ZO's double forward operations reduces unnecessary data movement, enhancing the fine-tuning efficacy. Additionally, our framework supports an innovative low-bit precision approach in AMP mode to streamline data exchanges between the CPU and GPU. Employing this approach allows us to fine-tune extraordinarily large models, such as the OPT-175B with more than 175 billion parameters, on a mere 18GB GPU--achievements beyond the reach of traditional methods. Moreover, our framework achieves these results with almost no additional time overhead and absolutely no accuracy loss compared to standard zeroth-order methods. ZO2's code has been open-sourced in https://github.com/liangyuwang/zo2.
We propose a cheaper alternative bilinear operator to matrix-multiplication in deep neural networks (DNNs). Unlike many stubborn attempts to accelerate MatMuls in DNN inference, this operator is supported by capabilities of existing GPU hardware, most notably NVIDIA TensorCores. To our knowledge, this is the first GPU-native acceleration technique which \emph{does not decrease} (in fact, increases) the number of trainable parameters of the network, mitigating the accuracy-loss of compression-based techniques. Hence, this operator is at the same time more expressive than MatMul, yet requires substantially \emph{fewer} FLOPs to evaluate. We term this new operator \emph{Strassen-Tile} (STL). The main idea behind STL$(X,W)$ is a \emph{local} change-of-basis (learnable encoder) on weights and activation \emph{tiles}, after which we perform batched \emph{elementwise} products between tiles, and a final decoding transformation (inspired by algebraic pipelines from fast matrix and polynomial multiplication). We compare STL against two benchmarks. The first one is SoTA T2T-ViT on Imagenet-1K. Here we show that replacing \emph{all} linear layers with STL and training from scratch, results in factor x2.7 reduction in FLOPs with a 0.5 \emph{accuracy improvement}. Our second speed-accuracy comparison benchmark for pretrained LLMs is the most practical GPU-acceleration technique, \twofour structured Sparsity. Finetuning TinyLlama \cite{tinyllama24} with STL layers on the Slim Pajama dataset, achieves similar accuracy to 2:4, with x2.2 FLOP speedup compared to x1.7 of the latter. Finally, we discuss a group-theoretic approach for discovering \emph{universal} encoders for STL, which could lead to fast \emph{black-box} acceleration via approximate matrix-multiplication (AMM).
In this study, we characterize GPU failures in Delta, the current large-scale AI system with over 600 petaflops of peak compute throughput. The system comprises GPU and non-GPU nodes with modern AI accelerators, such as NVIDIA A40, A100, and H100 GPUs. The study uses two and a half years of data on GPU errors. We evaluate the resilience of GPU hardware components to determine the vulnerability of different GPU components to failure and their impact on the GPU and node availability. We measure the key propagation paths in GPU hardware, GPU interconnect (NVLink), and GPU memory. Finally, we evaluate the impact of the observed GPU errors on user jobs. Our key findings are: (i) Contrary to common beliefs, GPU memory is over 30x more reliable than GPU hardware in terms of MTBE (mean time between errors). (ii) The newly introduced GSP (GPU System Processor) is the most vulnerable GPU hardware component. (iii) NVLink errors did not always lead to user job failure, and we attribute it to the underlying error detection and retry mechanisms employed. (iv) We show multiple examples of hardware errors originating from one of the key GPU hardware components, leading to application failure. (v) We project the impact of GPU node availability on larger scales with emulation and find that significant overprovisioning between 5-20% would be necessary to handle GPU failures. If GPU availability were improved to 99.9%, the overprovisioning would be reduced by 4x.
Timing optimization during the global placement of integrated circuits has been a significant focus for decades, yet it remains a complex, unresolved issue. Recent analytical methods typically use pin-level timing information to adjust net weights, which is fast and simple but neglects the path-based nature of the timing graph. The existing path-based methods, however, cannot balance the accuracy and efficiency due to the exponential growth of number of critical paths. In this work, we propose a GPU-accelerated timing-driven global placement framework, integrating accurate path-level information into the efficient DREAMPlace infrastructure. It optimizes the fine-grained pin-to-pin attraction objective and is facilitated by efficient critical path extraction. We also design a quadratic distance loss function specifically to align with the RC timing model. Experimental results demonstrate that our method significantly outperforms the current leading timing-driven placers, achieving an average improvement of 40.5% in total negative slack (TNS) and 8.3% in worst negative slack (WNS), as well as an improvement in half-perimeter wirelength (HPWL).
With the advancement of RNN models with linear complexity, the quadratic complexity challenge of transformers has the potential to be overcome. Notably, the emerging Mamba-2 has demonstrated competitive performance, bridging the gap between RNN models and transformers. However, due to sequential processing and vanishing gradients, RNN models struggle to capture long-range dependencies, limiting contextual understanding. This results in slow convergence, high resource demands, and poor performance on downstream understanding and complex reasoning tasks. In this work, we present a hybrid model MaTVLM by substituting a portion of the transformer decoder layers in a pre-trained VLM with Mamba-2 layers. Leveraging the inherent relationship between attention and Mamba-2, we initialize Mamba-2 with corresponding attention weights to accelerate convergence. Subsequently, we employ a single-stage distillation process, using the pre-trained VLM as the teacher model to transfer knowledge to the MaTVLM, further enhancing convergence speed and performance. Furthermore, we investigate the impact of differential distillation loss within our training framework. We evaluate the MaTVLM on multiple benchmarks, demonstrating competitive performance against the teacher model and existing VLMs while surpassing both Mamba-based VLMs and models of comparable parameter scales. Remarkably, the MaTVLM achieves up to 3.6x faster inference than the teacher model while reducing GPU memory consumption by 27.5%, all without compromising performance. Code and models are released at http://github.com/hustvl/MaTVLM.
Multi-head latent attention (MLA) is designed to optimize KV cache memory through low-rank key-value joint compression. Rather than caching keys and values separately, MLA stores their compressed latent representations, reducing memory overhead while maintaining the performance. While MLA improves memory efficiency without compromising language model accuracy, its major limitation lies in its integration during the pre-training phase, requiring models to be trained from scratch. This raises a key question: can we use MLA's benefits fully or partially in models that have already been pre-trained with different attention mechanisms? In this paper, we propose X-EcoMLA to deploy post training distillation to enable the upcycling of Transformer-based attention into an efficient hybrid MLA variant through lightweight post-training adaptation, bypassing the need for extensive pre-training. We demonstrate that leveraging the dark knowledge of a well-trained model can enhance training accuracy and enable extreme KV cache compression in MLA without compromising model performance. The experimental results show that our proposed method can effectively compress the KV cache while preserving the performance on the benchmarks; specifically, for Llama3.2-1B-Instruct baseline, a 6.4x compression achieves the same average score by using only 3.6B training tokens and 70 GPU hours on AMD MI300, whereas a 10.6x compression have less than 0.1\% average score drop with 7B training tokens and 140 GPU hours.
State-of-the-art transformer-based large multimodal models (LMMs) struggle to handle hour-long video inputs due to the quadratic complexity of the causal self-attention operations, leading to high computational costs during training and inference. Existing token compression-based methods reduce the number of video tokens but often incur information loss and remain inefficient for extremely long sequences. In this paper, we explore an orthogonal direction to build a hybrid Mamba-Transformer model (VAMBA) that employs Mamba-2 blocks to encode video tokens with linear complexity. Without any token reduction, VAMBA can encode more than 1024 frames (640$\times$360) on a single GPU, while transformer-based models can only encode 256 frames. On long video input, VAMBA achieves at least 50% reduction in GPU memory usage during training and inference, and nearly doubles the speed per training step compared to transformer-based LMMs. Our experimental results demonstrate that VAMBA improves accuracy by 4.3% on the challenging hour-long video understanding benchmark LVBench over prior efficient video LMMs, and maintains strong performance on a broad spectrum of long and short video understanding tasks.
Existing adaptation techniques typically require architectural modifications or added parameters, leading to high computational costs and complexity. We introduce Attention Projection Layer Adaptation (APLA), a simple approach to adapt vision transformers (ViTs) without altering the architecture or adding parameters. Through a systematic analysis, we find that the layer immediately after the attention mechanism is crucial for adaptation. By updating only this projection layer, or even just a random subset of this layer's weights, APLA achieves state-of-the-art performance while reducing GPU memory usage by up to 52.63% and training time by up to 43.0%, with no extra cost at inference. Across 46 datasets covering a variety of tasks including scene classification, medical imaging, satellite imaging, and fine-grained classification, APLA consistently outperforms 17 other leading adaptation methods, including full fine-tuning, on classification, segmentation, and detection tasks. The code is available at https://github.com/MoeinSorkhei/APLA.