PyTorchKR Papers

최신 논문

총 5,357편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2402.05406공식 코드 · Papers with Code

Everybody Prune Now: Structured Pruning of LLMs with only Forward Passes

Structured pruning is a promising approach to create smaller, faster LLMs. However, existing methods typically rely on backward passes, which can inflate memory requirements and compute costs. In this work we introduce Bonsai, a gradient-free structured pruning method that eliminates the need for backpropagation, significantly reducing memory requirements and compute costs while achieving state-of-the-art pruning performance. Bonsai uses forward-pass-only perturbative pruning to enable efficient compression of large models on a broader range of hardware configurations. Unlike existing structured pruning approaches, Bonsai not only achieves better compression with fewer resources, but also produces models that are twice as fast as those generated by semi-structured pruning. As a concrete demonstration, we use Bonsai to prune an 8B LLaMA-3 model to 50% sparsity on a single A6000 GPU -- a task infeasible with backprop-based methods, which require 2-3x memory. Our results show that removing backprop as a requirement not only enables pruning larger models on constrained hardware but can also lead to state-of-the-art efficiency and performance.

arXiv:2402.05291

Graph convolutional network as a fast statistical emulator for numerical ice sheet modeling

The Ice-sheet and Sea-level System Model (ISSM) provides numerical solutions for ice sheet dynamics using finite element and fine mesh adaption. However, considering ISSM is compatible only with central processing units (CPUs), it has limitations in economizing computational time to explore the linkage between climate forcings and ice dynamics. Although several deep learning emulators using graphic processing units (GPUs) have been proposed to accelerate ice sheet modeling, most of them rely on convolutional neural networks (CNNs) designed for regular grids. Since they are not appropriate for the irregular meshes of ISSM, we use a graph convolutional network (GCN) to replicate the adapted mesh structures of the ISSM. When applied to transient simulations of the Pine Island Glacier (PIG), Antarctica, the GCN successfully reproduces ice thickness and velocity with a correlation coefficient of approximately 0.997, outperforming non-graph models, including fully convolutional network (FCN) and multi-layer perceptron (MLP). Compared to the fixed-resolution approach of the FCN, the flexible-resolution structure of the GCN accurately captures detailed ice dynamics in fast-ice regions. By leveraging 60-100 times faster computational time of the GPU-based GCN emulator, we efficiently examine the impacts of basal melting rates on the ice sheet dynamics in the PIG.

arXiv:2402.05195

$λ$-ECLIPSE: Multi-Concept Personalized Text-to-Image Diffusion Models by Leveraging CLIP Latent Space

Despite the recent advances in personalized text-to-image (P-T2I) generative models, it remains challenging to perform finetuning-free multi-subject-driven T2I in a resource-efficient manner. Predominantly, contemporary approaches, involving the training of Hypernetworks and Multimodal Large Language Models (MLLMs), require heavy computing resources that range from 600 to 12300 GPU hours of training. These subject-driven T2I methods hinge on Latent Diffusion Models (LDMs), which facilitate T2I mapping through cross-attention layers. While LDMs offer distinct advantages, P-T2I methods' reliance on the latent space of these diffusion models significantly escalates resource demands, leading to inconsistent results and necessitating numerous iterations for a single desired image. In this paper, we present $\lambda$-ECLIPSE, an alternative prior-training strategy that works in the latent space of a pre-trained CLIP model without relying on the diffusion UNet models. $\lambda$-ECLIPSE leverages the image-text interleaved pre-training for fast and effective multi-subject-driven P-T2I. Through extensive experiments, we establish that $\lambda$-ECLIPSE surpasses existing baselines in composition alignment while preserving concept alignment performance, even with significantly lower resource utilization. $\lambda$-ECLIPSE performs multi-subject driven P-T2I with just 34M parameters and is trained on a mere 74 GPU hours. Additionally, $\lambda$-ECLIPSE demonstrates the unique ability to perform multi-concept interpolations.

arXiv:2402.05643공식 코드 · Papers with Code

Improving Token-Based World Models with Parallel Observation Prediction

Motivated by the success of Transformers when applied to sequences of discrete symbols, token-based world models (TBWMs) were recently proposed as sample-efficient methods. In TBWMs, the world model consumes agent experience as a language-like sequence of tokens, where each observation constitutes a sub-sequence. However, during imagination, the sequential token-by-token generation of next observations results in a severe bottleneck, leading to long training times, poor GPU utilization, and limited representations. To resolve this bottleneck, we devise a novel Parallel Observation Prediction (POP) mechanism. POP augments a Retentive Network (RetNet) with a novel forward mode tailored to our reinforcement learning setting. We incorporate POP in a novel TBWM agent named REM (Retentive Environment Model), showcasing a 15.4x faster imagination compared to prior TBWMs. REM attains superhuman performance on 12 out of 26 games of the Atari 100K benchmark, while training in less than 12 hours. Our code is available at \url{https://github.com/leor-c/REM}.

arXiv:2402.05396공식 코드 · Papers with Code

TASER: Temporal Adaptive Sampling for Fast and Accurate Dynamic Graph Representation Learning

Recently, Temporal Graph Neural Networks (TGNNs) have demonstrated state-of-the-art performance in various high-impact applications, including fraud detection and content recommendation. Despite the success of TGNNs, they are prone to the prevalent noise found in real-world dynamic graphs like time-deprecated links and skewed interaction distribution. The noise causes two critical issues that significantly compromise the accuracy of TGNNs: (1) models are supervised by inferior interactions, and (2) noisy input induces high variance in the aggregated messages. However, current TGNN denoising techniques do not consider the diverse and dynamic noise pattern of each node. In addition, they also suffer from the excessive mini-batch generation overheads caused by traversing more neighbors. We believe the remedy for fast and accurate TGNNs lies in temporal adaptive sampling. In this work, we propose TASER, the first adaptive sampling method for TGNNs optimized for accuracy, efficiency, and scalability. TASER adapts its mini-batch selection based on training dynamics and temporal neighbor selection based on the contextual, structural, and temporal properties of past interactions. To alleviate the bottleneck in mini-batch generation, TASER implements a pure GPU-based temporal neighbor finder and a dedicated GPU feature cache. We evaluate the performance of TASER using two state-of-the-art backbone TGNNs. On five popular datasets, TASER outperforms the corresponding baselines by an average of 2.3% in Mean Reciprocal Rank (MRR) while achieving an average of 5.1x speedup in training time.

arXiv:2402.04925

TP-Aware Dequantization

In this paper, we present a novel method that reduces model inference latency during distributed deployment of Large Language Models (LLMs). Our contribution is an optimized inference deployment scheme that address the current limitations of state-of-the-art quantization kernels when used in conjunction with Tensor Parallel (TP). Our method preserves data locality in GPU memory access patterns and exploits a priori knowledge of TP to reduce global communication. We demonstrate an up to 1.81x speedup over existing methods for Llama-70B and up to 1.78x speedup for IBM WatsonX's Granite-20B MLP layer problem sizes on A100 and H100 NVIDIA DGX Systems for a variety of TP settings.

arXiv:2402.04825공식 코드 · Papers with Code

Fast Timing-Conditioned Latent Audio Diffusion

Generating long-form 44.1kHz stereo audio from text prompts can be computationally demanding. Further, most previous works do not tackle that music and sound effects naturally vary in their duration. Our research focuses on the efficient generation of long-form, variable-length stereo music and sounds at 44.1kHz using text prompts with a generative model. Stable Audio is based on latent diffusion, with its latent defined by a fully-convolutional variational autoencoder. It is conditioned on text prompts as well as timing embeddings, allowing for fine control over both the content and length of the generated music and sounds. Stable Audio is capable of rendering stereo signals of up to 95 sec at 44.1kHz in 8 sec on an A100 GPU. Despite its compute efficiency and fast inference, it is one of the best in two public text-to-music and -audio benchmarks and, differently from state-of-the-art models, can generate music with structure and stereo sounds.

arXiv:2402.04554

BirdNeRF: Fast Neural Reconstruction of Large-Scale Scenes From Aerial Imagery

In this study, we introduce BirdNeRF, an adaptation of Neural Radiance Fields (NeRF) designed specifically for reconstructing large-scale scenes using aerial imagery. Unlike previous research focused on small-scale and object-centric NeRF reconstruction, our approach addresses multiple challenges, including (1) Addressing the issue of slow training and rendering associated with large models. (2) Meeting the computational demands necessitated by modeling a substantial number of images, requiring extensive resources such as high-performance GPUs. (3) Overcoming significant artifacts and low visual fidelity commonly observed in large-scale reconstruction tasks due to limited model capacity. Specifically, we present a novel bird-view pose-based spatial decomposition algorithm that decomposes a large aerial image set into multiple small sets with appropriately sized overlaps, allowing us to train individual NeRFs of sub-scene. This decomposition approach not only decouples rendering time from the scene size but also enables rendering to scale seamlessly to arbitrarily large environments. Moreover, it allows for per-block updates of the environment, enhancing the flexibility and adaptability of the reconstruction process. Additionally, we propose a projection-guided novel view re-rendering strategy, which aids in effectively utilizing the independently trained sub-scenes to generate superior rendering results. We evaluate our approach on existing datasets as well as against our own drone footage, improving reconstruction speed by 10x over classical photogrammetry software and 50x over state-of-the-art large-scale NeRF solution, on a single GPU with similar rendering quality.

arXiv:2402.04466

Towards Deterministic End-to-end Latency for Medical AI Systems in NVIDIA Holoscan

The introduction of AI and ML technologies into medical devices has revolutionized healthcare diagnostics and treatments. Medical device manufacturers are keen to maximize the advantages afforded by AI and ML by consolidating multiple applications onto a single platform. However, concurrent execution of several AI applications, each with its own visualization components, leads to unpredictable end-to-end latency, primarily due to GPU resource contentions. To mitigate this, manufacturers typically deploy separate workstations for distinct AI applications, thereby increasing financial, energy, and maintenance costs. This paper addresses these challenges within the context of NVIDIA's Holoscan platform, a real-time AI system for streaming sensor data and images. We propose a system design optimized for heterogeneous GPU workloads, encompassing both compute and graphics tasks. Our design leverages CUDA MPS for spatial partitioning of compute workloads and isolates compute and graphics processing onto separate GPUs. We demonstrate significant performance improvements across various end-to-end latency determinism metrics through empirical evaluation with real-world Holoscan medical device applications. For instance, the proposed design reduces maximum latency by 21-30% and improves latency distribution flatness by 17-25% for up to five concurrent endoscopy tool tracking AI applications, compared to a single-GPU baseline. Against a default multi-GPU setup, our optimizations decrease maximum latency by 35% for up to six concurrent applications by improving GPU utilization by 42%. This paper provides clear design insights for AI applications in the edge-computing domain including medical systems, where performance predictability of concurrent and heterogeneous GPU workloads is a critical requirement.

arXiv:2402.04291공식 코드 · Papers with Code

BiLLM: Pushing the Limit of Post-Training Quantization for LLMs

Pretrained large language models (LLMs) exhibit exceptional general language processing capabilities but come with significant demands on memory and computational resources. As a powerful compression technology, binarization can extremely reduce model weights to a mere 1 bit, lowering the expensive computation and memory requirements. However, existing quantization techniques fall short of maintaining LLM performance under ultra-low bit-widths. In response to this challenge, we present BiLLM, a groundbreaking 1-bit post-training quantization scheme tailored for pretrained LLMs. Based on the weight distribution of LLMs, BiLLM first identifies and structurally selects salient weights, and minimizes the compression loss through an effective binary residual approximation strategy. Moreover, considering the bell-shaped distribution of the non-salient weights, we propose an optimal splitting search to group and binarize them accurately. BiLLM achieving for the first time high-accuracy inference (e.g. 8.41 perplexity on LLaMA2-70B) with only 1.08-bit weights across various LLMs families and evaluation metrics, outperforms SOTA quantization methods of LLM by significant margins. Moreover, BiLLM enables the binarization process of the LLM with 7 billion weights within 0.5 hours on a single GPU, demonstrating satisfactory time efficiency. Our code is available at https://github.com/Aaronhuang-778/BiLLM.

arXiv:2402.05008공식 코드 · Papers with Code

EfficientViT-SAM: Accelerated Segment Anything Model Without Accuracy Loss

We present EfficientViT-SAM, a new family of accelerated segment anything models. We retain SAM's lightweight prompt encoder and mask decoder while replacing the heavy image encoder with EfficientViT. For the training, we begin with the knowledge distillation from the SAM-ViT-H image encoder to EfficientViT. Subsequently, we conduct end-to-end training on the SA-1B dataset. Benefiting from EfficientViT's efficiency and capacity, EfficientViT-SAM delivers 48.9x measured TensorRT speedup on A100 GPU over SAM-ViT-H without sacrificing performance. Our code and pre-trained models are released at https://github.com/mit-han-lab/efficientvit.

arXiv:2402.03548

Single-GPU GNN Systems: Traps and Pitfalls

The current graph neural network (GNN) systems have established a clear trend of not showing training accuracy results, and directly or indirectly relying on smaller datasets for evaluations majorly. Our in-depth analysis shows that it leads to a chain of pitfalls in the system design and evaluation process, questioning the practicality of many of the proposed system optimizations, and affecting conclusions and lessons learned. We analyze many single-GPU systems and show the fundamental impact of these pitfalls. We further develop hypotheses, recommendations, and evaluation methodologies, and provide future directions. Finally, a new reference system is developed to establish a new line of optimizations rooted in solving the system-design pitfalls efficiently and practically. The proposed design can productively be integrated into prior works, thereby truly advancing the state-of-the-art.

arXiv:2402.03460

Approximation Rates and VC-Dimension Bounds for (P)ReLU MLP Mixture of Experts

Mixture-of-Experts (MoEs) can scale up beyond traditional deep learning models by employing a routing strategy in which each input is processed by a single "expert" deep learning model. This strategy allows us to scale up the number of parameters defining the MoE while maintaining sparse activation, i.e., MoEs only load a small number of their total parameters into GPU VRAM for the forward pass depending on the input. In this paper, we provide an approximation and learning-theoretic analysis of mixtures of expert MLPs with (P)ReLU activation functions. We first prove that for every error level $\varepsilon>0$ and every Lipschitz function $f:[0,1]^n\to \mathbb{R}$, one can construct a MoMLP model (a Mixture-of-Experts comprising of (P)ReLU MLPs) which uniformly approximates $f$ to $\varepsilon$ accuracy over $[0,1]^n$, while only requiring networks of $\mathcal{O}(\varepsilon^{-1})$ parameters to be loaded in memory. Additionally, we show that MoMLPs can generalize since the entire MoMLP model has a (finite) VC dimension of $\tilde{O}(L\max\{nL,JW\})$, if there are $L$ experts and each expert has a depth and width of $J$ and $W$, respectively.

arXiv:2402.03908공식 코드 · Papers with Code

EscherNet: A Generative Model for Scalable View Synthesis

We introduce EscherNet, a multi-view conditioned diffusion model for view synthesis. EscherNet learns implicit and generative 3D representations coupled with a specialised camera positional encoding, allowing precise and continuous relative control of the camera transformation between an arbitrary number of reference and target views. EscherNet offers exceptional generality, flexibility, and scalability in view synthesis -- it can generate more than 100 consistent target views simultaneously on a single consumer-grade GPU, despite being trained with a fixed number of 3 reference views to 3 target views. As a result, EscherNet not only addresses zero-shot novel view synthesis, but also naturally unifies single- and multi-image 3D reconstruction, combining these diverse tasks into a single, cohesive framework. Our extensive experiments demonstrate that EscherNet achieves state-of-the-art performance in multiple benchmarks, even when compared to methods specifically tailored for each individual problem. This remarkable versatility opens up new directions for designing scalable neural architectures for 3D vision. Project page: https://kxhit.github.io/EscherNet.

arXiv:2402.03640

torchmSAT: A GPU-Accelerated Approximation To The Maximum Satisfiability Problem

The remarkable achievements of machine learning techniques in analyzing discrete structures have drawn significant attention towards their integration into combinatorial optimization algorithms. Typically, these methodologies improve existing solvers by injecting learned models within the solving loop to enhance the efficiency of the search process. In this work, we derive a single differentiable function capable of approximating solutions for the Maximum Satisfiability Problem (MaxSAT). Then, we present a novel neural network architecture to model our differentiable function, and progressively solve MaxSAT using backpropagation. This approach eliminates the need for labeled data or a neural network training phase, as the training process functions as the solving algorithm. Additionally, we leverage the computational power of GPUs to accelerate these computations. Experimental results on challenging MaxSAT instances show that our proposed methodology outperforms two existing MaxSAT solvers, and is on par with another in terms of solution cost, without necessitating any training or access to an underlying SAT solver. Given that numerous NP-hard problems can be reduced to MaxSAT, our novel technique paves the way for a new generation of solvers poised to benefit from neural network GPU acceleration.

arXiv:2402.04009

Low-rank Attention Side-Tuning for Parameter-Efficient Fine-Tuning

In finetuning a large pretrained model to downstream tasks, parameter-efficient fine-tuning (PEFT) methods can effectively finetune pretrained models with few trainable parameters, but suffer from high GPU memory consumption and slow training speed. Because learnable parameters from these methods are entangled with the pretrained model, gradients related to the frozen pretrained model's parameters have to be computed and stored during finetuning. We propose Low-rank Attention Side-Tuning (LAST), which disentangles the trainable module from the pretrained model by freezing not only parameters but also outputs of the pretrained network. LAST trains a side-network composed of only low-rank self-attention modules. By viewing the pretrained model as a frozen feature extractor, the side-network takes intermediate output from the pretrained model and focus on learning task-specific knowledge. We also show that LAST can be highly parallel across multiple optimization objectives, making it very efficient in downstream task adaptation, for example, in finding optimal hyperparameters. LAST outperforms previous state-of-the-art methods on VTAB-1K and other visual adaptation tasks with roughly only 30\% of GPU memory footprint and 60\% of training time compared to existing PEFT methods, but achieves significantly higher accuracy.

arXiv:2402.01975공식 코드 · Papers with Code

Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks

A molecule's 2D representation consists of its atoms, their attributes, and the molecule's covalent bonds. A 3D (geometric) representation of a molecule is called a conformer and consists of its atom types and Cartesian coordinates. Every conformer has a potential energy, and the lower this energy, the more likely it occurs in nature. Most existing machine learning methods for molecular property prediction consider either 2D molecular graphs or 3D conformer structure representations in isolation. Inspired by recent work on using ensembles of conformers in conjunction with 2D graph representations, we propose $\mathrm{E}$(3)-invariant molecular conformer aggregation networks. The method integrates a molecule's 2D representation with that of multiple of its conformers. Contrary to prior work, we propose a novel 2D-3D aggregation mechanism based on a differentiable solver for the Fused Gromov-Wasserstein Barycenter problem and the use of an efficient conformer generation method based on distance geometry. We show that the proposed aggregation mechanism is $\mathrm{E}$(3) invariant and propose an efficient GPU implementation. Moreover, we demonstrate that the aggregation mechanism helps to significantly outperform state-of-the-art molecule property prediction methods on established datasets.

arXiv:2402.01964공식 코드 · Papers with Code

Scalable and Efficient Temporal Graph Representation Learning via Forward Recent Sampling

Temporal graph representation learning (TGRL) is essential for modeling dynamic systems in real-world networks. However, traditional TGRL methods, despite their effectiveness, often face significant computational challenges and inference delays due to the inefficient sampling of temporal neighbors. Conventional sampling methods typically involve backtracking through the interaction history of each node. In this paper, we propose a novel TGRL framework, No-Looking-Back (NLB), which overcomes these challenges by introducing a forward recent sampling strategy. This strategy eliminates the need to backtrack through historical interactions by utilizing a GPU-executable, size-constrained hash table for each node. The hash table records a down-sampled set of recent interactions, enabling rapid query responses with minimal inference latency. The maintenance of this hash table is highly efficient, operating with $O(1)$ complexity. Fully compatible with GPU processing, NLB maximizes programmability, parallelism, and power efficiency. Empirical evaluations demonstrate that NLB not only matches or surpasses state-of-the-art methods in accuracy for tasks like link prediction and node classification across six real-world datasets but also achieves 1.32-4.40x faster training, 1.2-7.94x greater energy efficiency, and 1.63-12.95x lower inference latency compared to competitive baselines. The link to the code: https://github.com/Graph-COM/NLB.

arXiv:2402.01768

Enriched Physics-informed Neural Networks for Dynamic Poisson-Nernst-Planck Systems

This paper proposes a meshless deep learning algorithm, enriched physics-informed neural networks (EPINNs), to solve dynamic Poisson-Nernst-Planck (PNP) equations with strong coupling and nonlinear characteristics. The EPINNs takes the traditional physics-informed neural networks as the foundation framework, and adds the adaptive loss weight to balance the loss functions, which automatically assigns the weights of losses by updating the parameters in each iteration based on the maximum likelihood estimate. The resampling strategy is employed in the EPINNs to accelerate the convergence of loss function. Meanwhile, the GPU parallel computing technique is adopted to accelerate the solving process. Four examples are provided to demonstrate the validity and effectiveness of the proposed method. Numerical results indicate that the new method has better applicability than traditional numerical methods in solving such coupled nonlinear systems. More importantly, the EPINNs is more accurate, stable, and fast than the traditional physics-informed neural networks. This work provides a simple and high-performance numerical tool for addressing PNPs with arbitrary boundary shapes and boundary conditions.

arXiv:2402.01869공식 코드 · Papers with Code

InferCept: Efficient Intercept Support for Augmented Large Language Model Inference

Large language models are increasingly integrated with external environments, tools, and agents like ChatGPT plugins to extend their capability beyond language-centric tasks. However, today's LLM inference systems are designed for standalone LLMs. They treat each external interaction as the end of LLM generation and form a new request when the interaction finishes, causing unnecessary recomputation of already computed contexts, which accounts for 37-40% of total model forwarding time. This paper presents InferCept, the first LLM inference framework targeting augmented LLMs and supporting the efficient interception of LLM generation. InferCept minimizes the GPU resource waste caused by LLM interceptions and dedicates saved memory for serving more requests. InferCept improves the overall serving throughput by 1.6x-2x and completes 2x more requests per second compared to the state-of-the-art LLM inference systems.