PyTorchKR Papers

최신 논문

총 5,357편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2406.08334

ProTrain: Efficient LLM Training via Memory-Aware Techniques

It is extremely memory-hungry to train Large Language Models (LLM). To solve this problem, existing work exploits the combination of CPU and GPU for the training process, such as ZeRO-Offload. Such a technique largely democratizes billion-scale model training, making it possible to train with few consumer graphics cards. However, based on our observation, existing frameworks often provide coarse-grained memory management and require experienced experts in configuration tuning, leading to suboptimal hardware utilization and performance. This paper proposes ProTrain, a novel training system that intelligently balances memory usage and performance by coordinating memory, computation, and IO. ProTrain achieves adaptive memory management through Chunk-Based Model State Management and Block-Wise Activation Management, guided by a Memory-Aware Runtime Profiler without user intervention. ProTrain does not change the training algorithm and thus does not compromise accuracy. Experiments show that ProTrain improves training throughput by 1.43$\times$ to 2.71$\times$ compared to the SOTA training systems.

arXiv:2406.06571공식 코드 · Papers with Code

SUBLLM: A Novel Efficient Architecture with Token Sequence Subsampling for LLM

While Large Language Models (LLMs) have achieved remarkable success in various fields, the efficiency of training and inference remains a major challenge. To address this issue, we propose SUBLLM, short for Subsampling-Upsampling-Bypass Large Language Model, an innovative architecture that extends the core decoder-only framework by incorporating subsampling, upsampling, and bypass modules. The subsampling modules are responsible for shortening the sequence, while the upsampling modules restore the sequence length, and the bypass modules enhance convergence. In comparison to LLaMA, the proposed SUBLLM exhibits significant enhancements in both training and inference speeds as well as memory usage, while maintaining competitive few-shot performance. During training, SUBLLM increases speeds by 26% and cuts memory by 10GB per GPU. In inference, it boosts speeds by up to 37% and reduces memory by 1GB per GPU. The training and inference speeds can be enhanced by 34% and 52% respectively when the context window is expanded to 8192. Our code is available at https://github.com/XiaoMi/subllm.

arXiv:2406.06623공식 코드 · Papers with Code

Spectrum: Targeted Training on Signal to Noise Ratio

Efficiently post-training large language models remains a challenging task due to the vast computational resources required. We present Spectrum, a method that accelerates LLM training by selectively targeting layer modules based on their signal-to-noise ratio (SNR), and freezing the remaining modules. Our approach, which utilizes an algorithm to compute module SNRs prior to training, has shown to effectively match the performance of full fine-tuning while reducing GPU memory usage. Experiments comparing Spectrum to existing methods such as QLoRA demonstrate its effectiveness in terms of model quality and VRAM efficiency in distributed environments.

arXiv:2406.06858

FLUX: Fast Software-based Communication Overlap On GPUs Through Kernel Fusion

Large deep learning models have demonstrated strong ability to solve many tasks across a wide range of applications. Those large models typically require training and inference to be distributed. Tensor parallelism is a common technique partitioning computation of an operation or layer across devices to overcome the memory capacity limitation of a single processor, and/or to accelerate computation to meet a certain latency requirement. However, this kind of parallelism introduces additional communication that might contribute a significant portion of overall runtime. Thus limits scalability of this technique within a group of devices with high speed interconnects, such as GPUs with NVLinks in a node. This paper proposes a novel method, Flux, to significantly hide communication latencies with dependent computations for GPUs. Flux over-decomposes communication and computation operations into much finer-grained operations and further fuses them into a larger kernel to effectively hide communication without compromising kernel efficiency. Flux can potentially overlap up to 96% of communication given a fused kernel. Overall, it can achieve up to 1.24x speedups for training over Megatron-LM on a cluster of 128 GPUs with various GPU generations and interconnects, and up to 1.66x and 1.30x speedups for prefill and decoding inference over vLLM on a cluster with 8 GPUs with various GPU generations and interconnects.

arXiv:2406.07170

VoxNeuS: Enhancing Voxel-Based Neural Surface Reconstruction via Gradient Interpolation

Neural Surface Reconstruction learns a Signed Distance Field~(SDF) to reconstruct the 3D model from multi-view images. Previous works adopt voxel-based explicit representation to improve efficiency. However, they ignored the gradient instability of interpolation in the voxel grid, leading to degradation on convergence and smoothness. Besides, previous works entangled the optimization of geometry and radiance, which leads to the deformation of geometry to explain radiance, causing artifacts when reconstructing textured planes. In this work, we reveal that the instability of gradient comes from its discontinuity during trilinear interpolation, and propose to use the interpolated gradient instead of the original analytical gradient to eliminate the discontinuity. Based on gradient interpolation, we propose VoxNeuS, a lightweight surface reconstruction method for computational and memory efficient neural surface reconstruction. Thanks to the explicit representation, the gradient of regularization terms, i.e. Eikonal and curvature loss, are directly solved, avoiding computation and memory-access overhead. Further, VoxNeuS adopts a geometry-radiance disentangled architecture to handle the geometry deformation from radiance optimization. The experimental results show that VoxNeuS achieves better reconstruction quality than previous works. The entire training process takes 15 minutes and less than 3 GB of memory on a single 2080ti GPU.

arXiv:2406.07251공식 코드 · Papers with Code

Is One GPU Enough? Pushing Image Generation at Higher-Resolutions with Foundation Models

In this work, we introduce Pixelsmith, a zero-shot text-to-image generative framework to sample images at higher resolutions with a single GPU. We are the first to show that it is possible to scale the output of a pre-trained diffusion model by a factor of 1000, opening the road for gigapixel image generation at no additional cost. Our cascading method uses the image generated at the lowest resolution as a baseline to sample at higher resolutions. For the guidance, we introduce the Slider, a tunable mechanism that fuses the overall structure contained in the first-generated image with enhanced fine details. At each inference step, we denoise patches rather than the entire latent space, minimizing memory demands such that a single GPU can handle the process, regardless of the image's resolution. Our experimental results show that Pixelsmith not only achieves higher quality and diversity compared to existing techniques, but also reduces sampling time and artifacts. The code for our work is available at https://github.com/Thanos-DB/Pixelsmith.

arXiv:2406.07146

Benchmarking and Boosting Radiology Report Generation for 3D High-Resolution Medical Images

Automatic radiology report generation can significantly benefit the labor-intensive process of report writing by radiologists, especially for 3D radiographs like CT scans, which are crucial for broad clinical diagnostics yet underexplored compared to 2D radiographs. Existing methods often handle 3D volumes either slice-wise or with aggressive downsampling due to current GPU memory limitations, which results in a loss of the inherent 3D nature and critical details. To overcome these issues, we introduce a novel framework that efficiently and effectively generates radiology reports for high-resolution (HR) 3D volumes, based on large language models (LLMs). Specifically, our framework utilizes low-resolution (LR) visual tokens as queries to mine information from HR tokens, preserving detailed HR information while reducing computational costs by only processing HR informed LR visual queries. Further benefiting the field, we curate and release BIMCV-RG, a new dataset with 5,328 HR 3D volumes and paired reports, establishing the first benchmarks for report generation from 3D HR medical images. Our method consistently surpasses existing methods on this benchmark across three different settings: normal-resolution, high-resolution inputs, and zero-shot domain transfer, all at an acceptable computational cost, trainable on a single A100-80G.

arXiv:2406.05835공식 코드 · Papers with Code

Mamba YOLO: A Simple Baseline for Object Detection with State Space Model

Driven by the rapid development of deep learning technology, the YOLO series has set a new benchmark for real-time object detectors. Additionally, transformer-based structures have emerged as the most powerful solution in the field, greatly extending the model's receptive field and achieving significant performance improvements. However, this improvement comes at a cost as the quadratic complexity of the self-attentive mechanism increases the computational burden of the model. To address this problem, we introduce a simple yet effective baseline approach called Mamba YOLO. Our contributions are as follows: 1) We propose that the ODMamba backbone introduce a \textbf{S}tate \textbf{S}pace \textbf{M}odel (\textbf{SSM}) with linear complexity to address the quadratic complexity of self-attention. Unlike the other Transformer-base and SSM-base method, ODMamba is simple to train without pretraining. 2) For real-time requirement, we designed the macro structure of ODMamba, determined the optimal stage ratio and scaling size. 3) We design the RG Block that employs a multi-branch structure to model the channel dimensions, which addresses the possible limitations of SSM in sequence modeling, such as insufficient receptive fields and weak image localization. This design captures localized image dependencies more accurately and significantly. Extensive experiments on the publicly available COCO benchmark dataset show that Mamba YOLO achieves state-of-the-art performance compared to previous methods. Specifically, a tiny version of Mamba YOLO achieves a \textbf{7.5}\% improvement in mAP on a single 4090 GPU with an inference time of \textbf{1.5} ms. The pytorch code is available at: \url{https://github.com/HZAI-ZJNU/Mamba-YOLO}

arXiv:2406.05768공식 코드 · Papers with Code

TLCM: Training-efficient Latent Consistency Model for Image Generation with 2-8 Steps

Distilling latent diffusion models (LDMs) into ones that are fast to sample from is attracting growing research interest. However, the majority of existing methods face two critical challenges: (1) They hinge on long training using a huge volume of real data. (2) They routinely lead to quality degradation for generation, especially in text-image alignment. This paper proposes a novel training-efficient Latent Consistency Model (TLCM) to overcome these challenges. Our method first accelerates LDMs via data-free multistep latent consistency distillation (MLCD), and then data-free latent consistency distillation is proposed to efficiently guarantee the inter-segment consistency in MLCD. Furthermore, we introduce bags of techniques, e.g., distribution matching, adversarial learning, and preference learning, to enhance TLCM's performance at few-step inference without any real data. TLCM demonstrates a high level of flexibility by enabling adjustment of sampling steps within the range of 2 to 8 while still producing competitive outputs compared to full-step approaches. Notably, TLCM enjoys the data-free merit by employing synthetic data from the teacher for distillation. With just 70 training hours on an A100 GPU, a 3-step TLCM distilled from SDXL achieves an impressive CLIP Score of 33.68 and an Aesthetic Score of 5.97 on the MSCOCO-2017 5K benchmark, surpassing various accelerated models and even outperforming the teacher model in human preference metrics. We also demonstrate the versatility of TLCMs in applications including image style transfer, controllable generation, and Chinese-to-image generation.

arXiv:2406.06512

Merlin: A Vision Language Foundation Model for 3D Computed Tomography

Over 85 million computed tomography (CT) scans are performed annually in the US, of which approximately one quarter focus on the abdomen. Given the current radiologist shortage, there is a large impetus to use artificial intelligence to alleviate the burden of interpreting these complex imaging studies. Prior state-of-the-art approaches for automated medical image interpretation leverage vision language models (VLMs). However, current medical VLMs are generally limited to 2D images and short reports, and do not leverage electronic health record (EHR) data for supervision. We introduce Merlin - a 3D VLM that we train using paired CT scans (6+ million images from 15,331 CTs), EHR diagnosis codes (1.8+ million codes), and radiology reports (6+ million tokens). We evaluate Merlin on 6 task types and 752 individual tasks. The non-adapted (off-the-shelf) tasks include zero-shot findings classification (31 findings), phenotype classification (692 phenotypes), and zero-shot cross-modal retrieval (image to findings and image to impressions), while model adapted tasks include 5-year disease prediction (6 diseases), radiology report generation, and 3D semantic segmentation (20 organs). We perform internal validation on a test set of 5,137 CTs, and external validation on 7,000 clinical CTs and on two public CT datasets (VerSe, TotalSegmentator). Beyond these clinically-relevant evaluations, we assess the efficacy of various network architectures and training strategies to depict that Merlin has favorable performance to existing task-specific baselines. We derive data scaling laws to empirically assess training data needs for requisite downstream task performance. Furthermore, unlike conventional VLMs that require hundreds of GPUs for training, we perform all training on a single GPU.

arXiv:2406.06385공식 코드 · Papers with Code

Low-Rank Quantization-Aware Training for LLMs

Large language models (LLMs) are omnipresent, however their practical deployment is challenging due to their ever increasing computational and memory demands. Quantization is one of the most effective ways to make them more compute and memory efficient. Quantization-aware training (QAT) methods, generally produce the best quantized performance, however it comes at the cost of potentially long training time and excessive memory usage, making it impractical when applying for LLMs. Inspired by parameter-efficient fine-tuning (PEFT) and low-rank adaptation (LoRA) literature, we propose LR-QAT -- a lightweight and memory-efficient QAT algorithm for LLMs. LR-QAT employs several components to save memory without sacrificing predictive performance: (a) low-rank auxiliary weights that are aware of the quantization grid; (b) a downcasting operator using fixed-point or double-packed integers and (c) checkpointing. Unlike most related work, our method (i) is inference-efficient, leading to no additional overhead compared to traditional PTQ; (ii) can be seen as a general extended pretraining framework, meaning that the resulting model can still be utilized for any downstream task afterwards; (iii) can be applied across a wide range of quantization settings, such as different choices quantization granularity, activation quantization, and seamlessly combined with many PTQ techniques. We apply LR-QAT to LLaMA-1/2/3 and Mistral model families and validate its effectiveness on several downstream tasks. Our method outperforms common post-training quantization (PTQ) approaches and reaches the same model performance as full-model QAT at the fraction of its memory usage. Specifically, we can train a 7B LLM on a single consumer grade GPU with 24GB of memory. Our source code is available at https://github.com/qualcomm-ai-research/LR-QAT

arXiv:2406.06220공식 코드 · Papers with Code

Label-Looping: Highly Efficient Decoding for Transducers

This paper introduces a highly efficient greedy decoding algorithm for Transducer-based speech recognition models. We redesign the standard nested-loop design for RNN-T decoding, swapping loops over frames and labels: the outer loop iterates over labels, while the inner loop iterates over frames searching for the next non-blank symbol. Additionally, we represent partial hypotheses in a special structure using CUDA tensors, supporting parallelized hypotheses manipulations. Experiments show that the label-looping algorithm is up to 2.0X faster than conventional batched decoding when using batch size 32. It can be further combined with other compiler or GPU call-related techniques to achieve even more speedup. Our algorithm is general-purpose and can work with both conventional Transducers and Token-and-Duration Transducers. We open-source our implementation to benefit the research community.

arXiv:2406.04984공식 코드 · Papers with Code

MEFT: Memory-Efficient Fine-Tuning through Sparse Adapter

Parameter-Efficient Fine-tuning (PEFT) facilitates the fine-tuning of Large Language Models (LLMs) under limited resources. However, the fine-tuning performance with PEFT on complex, knowledge-intensive tasks is limited due to the constrained model capacity, which originates from the limited number of additional trainable parameters. To overcome this limitation, we introduce a novel mechanism that fine-tunes LLMs with adapters of larger size yet memory-efficient. This is achieved by leveraging the inherent activation sparsity in the Feed-Forward Networks (FFNs) of LLMs and utilizing the larger capacity of Central Processing Unit (CPU) memory compared to Graphics Processing Unit (GPU). We store and update the parameters of larger adapters on the CPU. Moreover, we employ a Mixture of Experts (MoE)-like architecture to mitigate unnecessary CPU computations and reduce the communication volume between the GPU and CPU. This is particularly beneficial over the limited bandwidth of PCI Express (PCIe). Our method can achieve fine-tuning results comparable to those obtained with larger memory capacities, even when operating under more limited resources such as a 24GB memory single GPU setup, with acceptable loss in training efficiency. Our codes are available at https://github.com/CURRENTF/MEFT.

arXiv:2406.04594

Enhancing Large-Scale AI Training Efficiency: The C4 Solution for Real-Time Anomaly Detection and Communication Optimization

The emergence of Large Language Models (LLMs) has necessitated the adoption of distributed training techniques, involving the deployment of thousands of GPUs to train a single model. Unfortunately, the efficiency of large-scale distributed training systems is often suboptimal due to the increased likelihood of hardware errors in high-end GPU products and the heightened risk of network traffic collisions. Moreover, any local hardware failure can disrupt training tasks, and the inability to swiftly identify faulty components leads to a significant waste of GPU resources. And, prolonged communication due to traffic collisions can substantially increase GPU waiting times. To address these challenges, we propose a communication-driven solution, namely the C4. The key insights of C4 are twofold. First, the load in distributed training exhibits homogeneous characteristics and is divided into iterations through periodic synchronization, therefore hardware anomalies would incur certain syndrome in collective communication. By leveraging this feature, C4 can rapidly identify the faulty components, swiftly isolate the anomaly, and restart the task, thereby avoiding resource wastage caused by delays in anomaly detection. Second, the predictable communication model of collective communication, involving a limited number of long-lived flows, allows C4 to efficiently execute traffic planning, substantially reducing bandwidth competition among these flows. The C4 has been extensively deployed across real-world production systems in a hyperscale cloud provider, yielding a significant improvement in system efficiency, from 30% to 45%. This enhancement is attributed to a 30% reduction in error-induced overhead and a 15% reduction in communication costs.

arXiv:2406.03923공식 코드 · Papers with Code

Latent Neural Operator for Solving Forward and Inverse PDE Problems

Neural operators effectively solve PDE problems from data without knowing the explicit equations, which learn the map from the input sequences of observed samples to the predicted values. Most existing works build the model in the original geometric space, leading to high computational costs when the number of sample points is large. We present the Latent Neural Operator (LNO) solving PDEs in the latent space. In particular, we first propose Physics-Cross-Attention (PhCA) transforming representation from the geometric space to the latent space, then learn the operator in the latent space, and finally recover the real-world geometric space via the inverse PhCA map. Our model retains flexibility that can decode values in any position not limited to locations defined in the training set, and therefore can naturally perform interpolation and extrapolation tasks particularly useful for inverse problems. Moreover, the proposed LNO improves both prediction accuracy and computational efficiency. Experiments show that LNO reduces the GPU memory by 50%, speeds up training 1.8 times, and reaches state-of-the-art accuracy on four out of six benchmarks for forward problems and a benchmark for inverse problem. Code is available at https://github.com/L-I-M-I-T/LatentNeuralOperator.

arXiv:2406.00291

Multi-Objective Neural Architecture Search by Learning Search Space Partitions

Deploying deep learning models requires taking into consideration neural network metrics such as model size, inference latency, and #FLOPs, aside from inference accuracy. This results in deep learning model designers leveraging multi-objective optimization to design effective deep neural networks in multiple criteria. However, applying multi-objective optimizations to neural architecture search (NAS) is nontrivial because NAS tasks usually have a huge search space, along with a non-negligible searching cost. This requires effective multi-objective search algorithms to alleviate the GPU costs. In this work, we implement a novel multi-objectives optimizer based on a recently proposed meta-algorithm called LaMOO on NAS tasks. In a nutshell, LaMOO speedups the search process by learning a model from observed samples to partition the search space and then focusing on promising regions likely to contain a subset of the Pareto frontier. Using LaMOO, we observe an improvement of more than 200% sample efficiency compared to Bayesian optimization and evolutionary-based multi-objective optimizers on different NAS datasets. For example, when combined with LaMOO, qEHVI achieves a 225% improvement in sample efficiency compared to using qEHVI alone in NasBench201. For real-world tasks, LaMOO achieves 97.36% accuracy with only 1.62M #Params on CIFAR10 in only 600 search samples. On ImageNet, our large model reaches 80.4% top-1 accuracy with only 522M #FLOPs.

arXiv:2406.00153

$μ$LO: Compute-Efficient Meta-Generalization of Learned Optimizers

Learned optimizers (LOs) can significantly reduce the wall-clock training time of neural networks, substantially reducing training costs. However, they often suffer from poor meta-generalization, especially when training networks larger than those seen during meta-training. To address this, we use the recently proposed Maximal Update Parametrization ($\mu$P), which allows zero-shot generalization of optimizer hyperparameters from smaller to larger models. We extend $\mu$P theory to learned optimizers, treating the meta-training problem as finding the learned optimizer under $\mu$P. Our evaluation shows that LOs meta-trained with $\mu$P substantially improve meta-generalization as compared to LOs trained under standard parametrization (SP). Notably, when applied to large-width models, our best $\mu$LO, trained for 103 GPU-hours, matches or exceeds the performance of VeLO, the largest publicly available learned optimizer, meta-trained with 4000 TPU-months of compute. Moreover, $\mu$LOs demonstrate better generalization than their SP counterparts to deeper networks and to much longer training horizons (25 times longer) than those seen during meta-training.

arXiv:2406.03791

Speed of Light Exact Greedy Decoding for RNN-T Speech Recognition Models on GPU

The vast majority of inference time for RNN Transducer (RNN-T) models today is spent on decoding. Current state-of-the-art RNN-T decoding implementations leave the GPU idle ~80% of the time. Leveraging a new CUDA 12.4 feature, CUDA graph conditional nodes, we present an exact GPU-based implementation of greedy decoding for RNN-T models that eliminates this idle time. Our optimizations speed up a 1.1 billion parameter RNN-T model end-to-end by a factor of 2.5x. This technique can applied to the "label looping" alternative greedy decoding algorithm as well, achieving 1.7x and 1.4x end-to-end speedups when applied to 1.1 billion parameter RNN-T and Token and Duration Transducer models respectively. This work enables a 1.1 billion parameter RNN-T model to run only 16% slower than a similarly sized CTC model, contradicting the common belief that RNN-T models are not suitable for high throughput inference. The implementation is available in NVIDIA NeMo.

arXiv:2406.03731공식 코드 · Papers with Code

Quality-Diversity with Limited Resources

Quality-Diversity (QD) algorithms have emerged as a powerful optimization paradigm with the aim of generating a set of high-quality and diverse solutions. To achieve such a challenging goal, QD algorithms require maintaining a large archive and a large population in each iteration, which brings two main issues, sample and resource efficiency. Most advanced QD algorithms focus on improving the sample efficiency, while the resource efficiency is overlooked to some extent. Particularly, the resource overhead during the training process has not been touched yet, hindering the wider application of QD algorithms. In this paper, we highlight this important research question, i.e., how to efficiently train QD algorithms with limited resources, and propose a novel and effective method called RefQD to address it. RefQD decomposes a neural network into representation and decision parts, and shares the representation part with all decision parts in the archive to reduce the resource overhead. It also employs a series of strategies to address the mismatch issue between the old decision parts and the newly updated representation part. Experiments on different types of tasks from small to large resource consumption demonstrate the excellent performance of RefQD: it not only uses significantly fewer resources (e.g., 16\% GPU memories on QDax and 3.7\% on Atari) but also achieves comparable or better performance compared to sample-efficient QD algorithms. Our code is available at \url{https://github.com/lamda-bbo/RefQD}.

arXiv:2406.04343공식 코드 · Papers with Code

Flash3D: Feed-Forward Generalisable 3D Scene Reconstruction from a Single Image

In this paper, we propose Flash3D, a method for scene reconstruction and novel view synthesis from a single image which is both very generalisable and efficient. For generalisability, we start from a "foundation" model for monocular depth estimation and extend it to a full 3D shape and appearance reconstructor. For efficiency, we base this extension on feed-forward Gaussian Splatting. Specifically, we predict a first layer of 3D Gaussians at the predicted depth, and then add additional layers of Gaussians that are offset in space, allowing the model to complete the reconstruction behind occlusions and truncations. Flash3D is very efficient, trainable on a single GPU in a day, and thus accessible to most researchers. It achieves state-of-the-art results when trained and tested on RealEstate10k. When transferred to unseen datasets like NYU it outperforms competitors by a large margin. More impressively, when transferred to KITTI, Flash3D achieves better PSNR than methods trained specifically on that dataset. In some instances, it even outperforms recent methods that use multiple views as input. Code, models, demo, and more results are available at https://www.robots.ox.ac.uk/~vgg/research/flash3d/.