Large Language Models (LLMs) demonstrate remarkable emergent abilities across various tasks, yet fall short of complex reasoning and planning tasks. The tree-search-based reasoning methods address this by surpassing the capabilities of chain-of-thought prompting, encouraging exploration of intermediate steps. However, such methods introduce significant inference latency due to the systematic exploration and evaluation of multiple thought paths. This paper introduces SeeD, a novel and efficient inference framework to optimize runtime speed and GPU memory management concurrently. By employing a scheduled speculative execution, SeeD efficiently handles multiple iterations for the thought generation and the state evaluation, leveraging a rounds-scheduled strategy to manage draft model dispatching. Extensive experimental evaluations on three reasoning datasets demonstrate superior speedup performance of SeeD, providing a viable path for batched inference in training-free speculative decoding.
Reconstructing 3D hand-face interactions with deformations from a single image is a challenging yet crucial task with broad applications in AR, VR, and gaming. The challenges stem from self-occlusions during single-view hand-face interactions, diverse spatial relationships between hands and face, complex deformations, and the ambiguity of the single-view setting. The first and only method for hand-face interaction recovery, Decaf, introduces a global fitting optimization guided by contact and deformation estimation networks trained on studio-collected data with 3D annotations. However, Decaf suffers from a time-consuming optimization process and limited generalization capability due to its reliance on 3D annotations of hand-face interaction data. To address these issues, we present DICE, the first end-to-end method for Deformation-aware hand-face Interaction reCovEry from a single image. DICE estimates the poses of hands and faces, contacts, and deformations simultaneously using a Transformer-based architecture. It features disentangling the regression of local deformation fields and global mesh vertex locations into two network branches, enhancing deformation and contact estimation for precise and robust hand-face mesh recovery. To improve generalizability, we propose a weakly-supervised training approach that augments the training set using in-the-wild images without 3D ground-truth annotations, employing the depths of 2D keypoints estimated by off-the-shelf models and adversarial priors of poses for supervision. Our experiments demonstrate that DICE achieves state-of-the-art performance on a standard benchmark and in-the-wild data in terms of accuracy and physical plausibility. Additionally, our method operates at an interactive rate (20 fps) on an Nvidia 4090 GPU, whereas Decaf requires more than 15 seconds for a single image. Our code will be publicly available upon publication.
3D Gaussian Splatting (3DGS) is increasingly popular for 3D reconstruction due to its superior visual quality and rendering speed. However, 3DGS training currently occurs on a single GPU, limiting its ability to handle high-resolution and large-scale 3D reconstruction tasks due to memory constraints. We introduce Grendel, a distributed system designed to partition 3DGS parameters and parallelize computation across multiple GPUs. As each Gaussian affects a small, dynamic subset of rendered pixels, Grendel employs sparse all-to-all communication to transfer the necessary Gaussians to pixel partitions and performs dynamic load balancing. Unlike existing 3DGS systems that train using one camera view image at a time, Grendel supports batched training with multiple views. We explore various optimization hyperparameter scaling strategies and find that a simple sqrt(batch size) scaling rule is highly effective. Evaluations using large-scale, high-resolution scenes show that Grendel enhances rendering quality by scaling up 3DGS parameters across multiple GPUs. On the Rubble dataset, we achieve a test PSNR of 27.28 by distributing 40.4 million Gaussians across 16 GPUs, compared to a PSNR of 26.28 using 11.2 million Gaussians on a single GPU. Grendel is an open-source project available at: https://github.com/nyu-systems/Grendel-GS
This paper introduces ConStyle v2, a strong plug-and-play prompter designed to output clean visual prompts and assist U-Net Image Restoration models in handling multiple degradations. The joint training process of IRConStyle, an Image Restoration framework consisting of ConStyle and a general restoration network, is divided into two stages: first, pre-training ConStyle alone, and then freezing its weights to guide the training of the general restoration network. Three improvements are proposed in the pre-training stage to train ConStyle: unsupervised pre-training, adding a pretext task (i.e. classification), and adopting knowledge distillation. Without bells and whistles, we can get ConStyle v2, a strong prompter for all-in-one Image Restoration, in less than two GPU days and doesn't require any fine-tuning. Extensive experiments on Restormer (transformer-based), NAFNet (CNN-based), MAXIM-1S (MLP-based), and a vanilla CNN network demonstrate that ConStyle v2 can enhance any U-Net style Image Restoration models to all-in-one Image Restoration models. Furthermore, models guided by the well-trained ConStyle v2 exhibit superior performance in some specific degradation compared to ConStyle.
This article introduces the structure flow field; a flow field that can provide high-speed robo-centric motion information for motion control of highly dynamic robotic devices and autonomous vehicles. Structure flow is the angular 3D velocity of the scene at a given pixel. We show that structure flow posses an elegant evolution model in the form of a Partial Differential Equation (PDE) that enables us to create dense flow predictions forward in time. We exploit this structure to design a predictor-update algorithm to compute structure flow in real time using image and depth measurements. The prediction stage takes the previous estimate of the structure flow and propagates it forward in time using a numerical implementation of the structure flow PDE. The predicted flow is then updated using new image and depth data. The algorithm runs up to 600 Hz on a Desktop GPU machine for 512x512 images with flow values up to 8 pixels. We provide ground truth validation on high-speed synthetic image sequences as well as results on real-life video on driving scenarios.
We introduce the Overcooked Generalisation Challenge (OGC) - the first benchmark to study agents' zero-shot cooperation abilities when faced with novel partners and levels in the Overcooked-AI environment. This perspective starkly contrasts a large body of previous work that has trained and evaluated cooperating agents only on the same level, failing to capture generalisation abilities required for real-world human-AI cooperation. Our challenge interfaces with state-of-the-art dual curriculum design (DCD) methods to generate auto-curricula for training general agents in Overcooked. It is the first cooperative multi-agent environment specially designed for DCD methods and, consequently, the first benchmarked with state-of-the-art methods. It is fully GPU-accelerated, built on the DCD benchmark suite minimax, and freely available under an open-source license: https://git.hcics.simtech.uni-stuttgart.de/public-projects/OGC. We show that current DCD algorithms struggle to produce useful policies in this novel challenge, even if combined with recent network architectures that were designed for scalability and generalisability. The OGC pushes the boundaries of real-world human-AI cooperation by enabling the research community to study the impact of generalisation on cooperating agents.
Training large language models (LLMs) for pretraining or adapting to new tasks and domains has become increasingly critical as their applications expand. However, as the model and the data sizes grow, the training process presents significant memory challenges, often requiring a prohibitive amount of GPU memory that may not be readily available. Existing methods such as low-rank adaptation (LoRA) add trainable low-rank matrix factorizations, altering the training dynamics and limiting the model's parameter search to a low-rank subspace. GaLore, a more recent method, employs Gradient Low-Rank Projection to reduce the memory footprint, in the full parameter training setting. However GaLore can only be applied to a subset of the LLM layers that satisfy the "reversibility" property, thus limiting their applicability. In response to these challenges, we introduce BlockLLM, an approach inspired by block coordinate descent. Our method carefully selects and updates a very small subset of the trainable parameters without altering any part of its architecture and training procedure. BlockLLM achieves state-of-the-art performance in both finetuning and pretraining tasks, while reducing the memory footprint of the underlying optimization process. Our experiments demonstrate that fine-tuning with only less than 5% of the parameters, BlockLLM achieves state-of-the-art perplexity scores on the GLUE benchmarks. On Llama model pretrained on C4 dataset, BlockLLM is able to train with significantly less memory than the state-of-the-art, while still maintaining competitive performance.
Large language model (LLM) training and finetuning are often bottlenecked by limited GPU memory. While existing projection-based optimization methods address this by projecting gradients into a lower-dimensional subspace to reduce optimizer state memory, they typically rely on dense projection matrices, which can introduce computational and memory overheads. In this work, we propose Grass (GRAdient Stuctured Sparsification), a novel approach that leverages sparse projections to transform gradients into structured sparse updates. This design not only significantly reduces memory usage for optimizer states but also minimizes gradient memory footprint, computation, and communication costs, leading to substantial throughput improvements. Extensive experiments on pretraining and finetuning tasks demonstrate that Grass achieves competitive performance to full-rank training and existing projection-based methods. Notably, Grass enables half-precision pretraining of a 13B parameter LLaMA model on a single 40GB A100 GPU--a feat infeasible for previous methods--and yields up to a $2\times$ throughput improvement on an 8-GPU system. Code can be found at https://github.com/aashiqmuhamed/GRASS .
Neural implicit functions have demonstrated significant importance in various areas such as computer vision, graphics. Their advantages include the ability to represent complex shapes and scenes with high fidelity, smooth interpolation capabilities, and continuous representations. Despite these benefits, the development and analysis of implicit functions have been limited by the lack of comprehensive datasets and the substantial computational resources required for their implementation and evaluation. To address these challenges, we introduce "Implicit-Zoo": a large-scale dataset requiring thousands of GPU training days designed to facilitate research and development in this field. Our dataset includes diverse 2D and 3D scenes, such as CIFAR-10, ImageNet-1K, and Cityscapes for 2D image tasks, and the OmniObject3D dataset for 3D vision tasks. We ensure high quality through strict checks, refining or filtering out low-quality data. Using Implicit-Zoo, we showcase two immediate benefits as it enables to: (1) learn token locations for transformer models; (2) directly regress 3D cameras poses of 2D images with respect to NeRF models. This in turn leads to an improved performance in all three task of image classification, semantic segmentation, and 3D pose regression, thereby unlocking new avenues for research.
Deep neural networks (DNNs) continue to grow rapidly in size, making them infeasible to train on a single device. Pipeline parallelism is commonly used in existing DNN systems to support large-scale DNN training by partitioning a DNN into multiple stages, which concurrently perform DNN training for different micro-batches in a pipeline fashion. However, existing pipeline-parallel approaches only consider sequential pipeline stages and thus ignore the topology of a DNN, resulting in missed model-parallel opportunities. This paper presents graph pipeline parallelism (GPP), a new pipeline-parallel scheme that partitions a DNN into pipeline stages whose dependencies are identified by a directed acyclic graph. GPP generalizes existing sequential pipeline parallelism and preserves the inherent topology of a DNN to enable concurrent execution of computationally-independent operators, resulting in reduced memory requirement and improved GPU performance. In addition, we develop GraphPipe, a distributed system that exploits GPP strategies to enable performant and scalable DNN training. GraphPipe partitions a DNN into a graph of stages, optimizes micro-batch schedules for these stages, and parallelizes DNN training using the discovered GPP strategies. Evaluation on a variety of DNNs shows that GraphPipe outperforms existing pipeline-parallel systems such as PipeDream and Piper by up to 1.6X. GraphPipe also reduces the search time by 9-21X compared to PipeDream and Piper.
The increasing deployment of artificial intelligence (AI) for critical decision-making amplifies the necessity for trustworthy AI, where uncertainty estimation plays a pivotal role in ensuring trustworthiness. Dropout-based Bayesian Neural Networks (BayesNNs) are prominent in this field, offering reliable uncertainty estimates. Despite their effectiveness, existing dropout-based BayesNNs typically employ a uniform dropout design across different layers, leading to suboptimal performance. Moreover, as diverse applications require tailored dropout strategies for optimal performance, manually optimizing dropout configurations for various applications is both error-prone and labor-intensive. To address these challenges, this paper proposes a novel neural dropout search framework that automatically optimizes both the dropout-based BayesNNs and their hardware implementations on FPGA. We leverage one-shot supernet training with an evolutionary algorithm for efficient dropout optimization. A layer-wise dropout search space is introduced to enable the automatic design of dropout-based BayesNNs with heterogeneous dropout configurations. Extensive experiments demonstrate that our proposed framework can effectively find design configurations on the Pareto frontier. Compared to manually-designed dropout-based BayesNNs on GPU, our search approach produces FPGA designs that can achieve up to 33X higher energy efficiency. Compared to state-of-the-art FPGA designs of BayesNN, the solutions from our approach can achieve higher algorithmic performance and energy efficiency.
Deep neural networks (DNNs) typically employ an end-to-end (E2E) training paradigm which presents several challenges, including high GPU memory consumption, inefficiency, and difficulties in model parallelization during training. Recent research has sought to address these issues, with one promising approach being local learning. This method involves partitioning the backbone network into gradient-isolated modules and manually designing auxiliary networks to train these local modules. Existing methods often neglect the interaction of information between local modules, leading to myopic issues and a performance gap compared to E2E training. To address these limitations, we propose the Multilaminar Leap Augmented Auxiliary Network (MLAAN). Specifically, MLAAN comprises Multilaminar Local Modules (MLM) and Leap Augmented Modules (LAM). MLM captures both local and global features through independent and cascaded auxiliary networks, alleviating performance issues caused by insufficient global features. However, overly simplistic auxiliary networks can impede MLM's ability to capture global information. To address this, we further design LAM, an enhanced auxiliary network that uses the Exponential Moving Average (EMA) method to facilitate information exchange between local modules, thereby mitigating the shortsightedness resulting from inadequate interaction. The synergy between MLM and LAM has demonstrated excellent performance. Our experiments on the CIFAR-10, STL-10, SVHN, and ImageNet datasets show that MLAAN can be seamlessly integrated into existing local learning frameworks, significantly enhancing their performance and even surpassing end-to-end (E2E) training methods, while also reducing GPU memory consumption.
With the increasing prevalence of autonomous vehicles, it is essential for computer vision algorithms to accurately assess road features in real-time. This study explores the LaneSegNet architecture, a new approach to lane topology prediction which integrates topological information with lane-line data to provide a more contextual understanding of road environments. The LaneSegNet architecture includes a feature extractor, lane encoder, lane decoder, and prediction head, leveraging components from ResNet-50, BEVFormer, and various attention mechanisms. We experimented with optimizations to the LaneSegNet architecture through feature extractor modification and transformer encoder-decoder stack modification. We found that modifying the encoder and decoder stacks offered an interesting tradeoff between training time and prediction accuracy, with certain combinations showing promising results. Our implementation, trained on a single NVIDIA Tesla A100 GPU, found that a 2:4 ratio reduced training time by 22.3% with only a 7.1% drop in mean average precision, while a 4:8 ratio increased training time by only 11.1% but improved mean average precision by a significant 23.7%. These results indicate that strategic hyperparameter tuning can yield substantial improvements depending on the resources of the user. This study provides valuable insights for optimizing LaneSegNet according to available computation power, making it more accessible for users with limited resources and increasing the capabilities for users with more powerful resources.
Diffusion models have recently achieved remarkable results for video generation. Despite the encouraging performances, the generated videos are typically constrained to a small number of frames, resulting in clips lasting merely a few seconds. The primary challenges in producing longer videos include the substantial memory requirements and the extended processing time required on a single GPU. A straightforward solution would be to split the workload across multiple GPUs, which, however, leads to two issues: (1) ensuring all GPUs communicate effectively to share timing and context information, and (2) modifying existing video diffusion models, which are usually trained on short sequences, to create longer videos without additional training. To tackle these, in this paper we introduce Video-Infinity, a distributed inference pipeline that enables parallel processing across multiple GPUs for long-form video generation. Specifically, we propose two coherent mechanisms: Clip parallelism and Dual-scope attention. Clip parallelism optimizes the gathering and sharing of context information across GPUs which minimizes communication overhead, while Dual-scope attention modulates the temporal self-attention to balance local and global contexts efficiently across the devices. Together, the two mechanisms join forces to distribute the workload and enable the fast generation of long videos. Under an 8 x Nvidia 6000 Ada GPU (48G) setup, our method generates videos up to 2,300 frames in approximately 5 minutes, enabling long video generation at a speed 100 times faster than the prior methods.
Training recommendation systems (RecSys) faces several challenges as it requires the "data preprocessing" stage to preprocess an ample amount of raw data and feed them to the GPU for training in a seamless manner. To sustain high training throughput, state-of-the-art solutions reserve a large fleet of CPU servers for preprocessing which incurs substantial deployment cost and power consumption. Our characterization reveals that prior CPU-centric preprocessing is bottlenecked on feature generation and feature normalization operations as it fails to reap out the abundant inter-/intra-feature parallelism in RecSys preprocessing. PreSto is a storage-centric preprocessing system leveraging In-Storage Processing (ISP), which offloads the bottlenecked preprocessing operations to our ISP units. We show that PreSto outperforms the baseline CPU-centric system with a $9.6\times$ speedup in end-to-end preprocessing time, $4.3\times$ enhancement in cost-efficiency, and $11.3\times$ improvement in energyefficiency on average for production-scale RecSys preprocessing.
Reliable uncertainty estimation plays a crucial role in various safety-critical applications such as medical diagnosis and autonomous driving. In recent years, Bayesian neural networks (BayesNNs) have gained substantial research and industrial interests due to their capability to make accurate predictions with reliable uncertainty estimation. However, the algorithmic complexity and the resulting hardware performance of BayesNNs hinder their adoption in real-life applications. To bridge this gap, this paper proposes an algorithm and hardware co-design framework that can generate field-programmable gate array (FPGA)-based accelerators for efficient BayesNNs. At the algorithm level, we propose novel multi-exit dropout-based BayesNNs with reduced computational and memory overheads while achieving high accuracy and quality of uncertainty estimation. At the hardware level, this paper introduces a transformation framework that can generate FPGA-based accelerators for the proposed efficient multi-exit BayesNNs. Several optimization techniques such as the mix of spatial and temporal mappings are introduced to reduce resource consumption and improve the overall hardware performance. Comprehensive experiments demonstrate that our approach can achieve higher energy efficiency compared to CPU, GPU, and other state-of-the-art hardware implementations. To support the future development of this research, we have open-sourced our code at: https://github.com/os-hxfan/MCME_FPGA_Acc.git
In this work, we introduce the Geometry-Aware Large Reconstruction Model (GeoLRM), an approach which can predict high-quality assets with 512k Gaussians and 21 input images in only 11 GB GPU memory. Previous works neglect the inherent sparsity of 3D structure and do not utilize explicit geometric relationships between 3D and 2D images. This limits these methods to a low-resolution representation and makes it difficult to scale up to the dense views for better quality. GeoLRM tackles these issues by incorporating a novel 3D-aware transformer structure that directly processes 3D points and uses deformable cross-attention mechanisms to effectively integrate image features into 3D representations. We implement this solution through a two-stage pipeline: initially, a lightweight proposal network generates a sparse set of 3D anchor points from the posed image inputs; subsequently, a specialized reconstruction transformer refines the geometry and retrieves textural details. Extensive experimental results demonstrate that GeoLRM significantly outperforms existing models, especially for dense view inputs. We also demonstrate the practical applicability of our model with 3D generation tasks, showcasing its versatility and potential for broader adoption in real-world applications. The project page: https://linshan-bin.github.io/GeoLRM/.
Sparse attention can effectively mitigate the significant memory and throughput demands of Large Language Models (LLMs) in long contexts. Existing methods typically employ a uniform sparse attention mask, applying the same sparse pattern across different attention heads and input lengths. However, this uniform approach fails to capture the diverse attention patterns inherent in LLMs, ignoring their distinct accuracy-latency trade-offs. To address this challenge, we propose the Mixture of Attention (MoA), which automatically tailors distinct sparse attention configurations to different heads and layers. MoA constructs and navigates a search space of various attention patterns and their scaling rules relative to input sequence lengths. It profiles the model, evaluates potential configurations, and pinpoints the optimal sparse attention compression plan. MoA adapts to varying input sizes, revealing that some attention heads expand their focus to accommodate longer sequences, while other heads consistently concentrate on fixed-length local contexts. Experiments show that MoA increases the effective context length by $3.9\times$ with the same average attention span, boosting retrieval accuracy by $1.5-7.1\times$ over the uniform-attention baseline across Vicuna-{7B,13B}, and Llama3-{8B,70B} models. Moreover, MoA narrows the capability gaps between sparse and dense models, reducing the maximum relative performance drop from $9\%-36\%$ to within $5\%$ across two long-context understanding benchmarks. MoA achieves a $1.2-1.4\times$ GPU memory reduction, boosting decode throughput by $6.6-8.2\times$ and $1.7-1.9\times$ compared to FlashAttention2 and vLLM, with minimal impact on performance. Our code is available at \url{https://github.com/thu-nics/MoA}.
DC Optimal Power Flow (DCOPF) is a key operational tool for power system operators, and it is embedded as a subproblem in many challenging optimization problems (e.g., line switching). However, traditional CPU-based solve routines (e.g., simplex) have saturated in speed and are hard to parallelize. This paper focuses on solving DCOPF problems using gradient-based routines on Graphics Processing Units (GPUs), which have massive parallelization capability. To formulate these problems, we pose a Lagrange dual associated with DCOPF (linear and quadratic cost curves), and then we explicitly solve the inner (primal) minimization problem with a dual norm. The resulting dual problem can be efficiently iterated using projected gradient ascent. After solving the dual problem on both CPUs and GPUs to find tight lower bounds, we benchmark against Gurobi and MOSEK, comparing convergence speed and tightness on the IEEE 2000, 4601, and 10000 bus systems. We provide reliable and tight lower bounds for these problems with, at best, 5.4x speedup over a conventional solver.
The ability to train generative models that produce realistic, safe and useful tabular data is essential for data privacy, imputation, oversampling, explainability or simulation. However, generating tabular data is not straightforward due to its heterogeneity, non-smooth distributions, complex dependencies and imbalanced categorical features. Although diverse methods have been proposed in the literature, there is a need for a unified evaluation, under the same conditions, on a variety of datasets. This study addresses this need by fully considering the optimization of: hyperparameters, feature encodings, and architectures. We investigate the impact of dataset-specific tuning on five recent model families for tabular data generation through an extensive benchmark on 16 datasets. These datasets vary in terms of size (an average of 80,000 rows), data types, and domains. We also propose a reduced search space for each model that allows for quick optimization, achieving nearly equivalent performance at a significantly lower cost. Our benchmark demonstrates that, for most models, large-scale dataset-specific tuning substantially improves performance compared to the original configurations. Furthermore, we confirm that diffusion-based models generally outperform other models on tabular data. However, this advantage is not significant when the entire tuning and training process is restricted to the same GPU budget.