OVERLORD: Ultimate Scaling of DataLoader for Multi-Source Large Foundation Model Training

• URL: http://arxiv.org/abs/2504.09844v1
• Date: Mon, 14 Apr 2025 03:31:22 GMT
• Title: OVERLORD: Ultimate Scaling of DataLoader for Multi-Source Large Foundation Model Training
• Authors: Juntao Zhao, Qi Lu, Wei Jia, Borui Wan, Lei Zuo, Junda Feng, Jianyu Jiang, Yangrui Chen, Shuaishuai Cao, Jialing He, Kaihua Jiang, Yuanzhe Hu, Yanghua Peng, Haibin Lin, Xin Liu, Chuan Wu,
• Abstract summary: We present OVERLORD, an industrial-grade distributed data loading architecture with three innovations.<n> OVERLORD achieves: (1) 4.5x end-to-end training throughput improvement; (2) a minimum 3.6x reduction in CPU memory usage.
• Score: 17.215899004049778
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Modern frameworks for training large foundation models (LFMs) employ data loaders in a data parallel paradigm. While this design offers implementation simplicity, it introduces two fundamental challenges. First, due to the quadratic computational complexity of the attention operator, the non-uniform sample distribution over data-parallel ranks leads to a significant workload imbalance among loaders, which degrades the training efficiency. This paradigm also impedes the implementation of data mixing algorithms (e.g., curriculum learning) over different datasets. Second, to acquire a broad range of capability, LFMs training ingests data from diverse sources, each with distinct file access states. Colocating massive datasets within loader instances can easily exceed local pod memory capacity. Additionally, heavy sources with higher transformation latency require larger worker pools, further exacerbating memory consumption. We present OVERLORD, an industrial-grade distributed data loading architecture with three innovations: (1) A centralized and declarative data plane, which facilitates elastic data orchestration strategy, such as long-short context, multimodal, and curriculum learning; (2) Disaggregated multisource preprocessing through role-specific actors, i.e., Source Loaders and Data Constructors, leveraging autoscaling for Source Loaders towards heterogeneous and evolving source preprocessing cost; (3) Shadow Loaders with differential checkpointing for uninterrupted fault recovery. Deployed on production clusters scaling to multi-thousand GPU, OVERLORD achieves: (1) 4.5x end-to-end training throughput improvement, (2) a minimum 3.6x reduction in CPU memory usage, with further improvements to be added in later experiments.

Related papers

• StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation [55.75008325187133]
  Reinforcement learning (RL) has become the core post-training technique for large language models (LLMs) StreamRL is designed with disaggregation from first principles to address two types of performance bottlenecks. Experiments show that StreamRL improves throughput by up to 2.66x compared to existing state-of-the-art systems.
  arXiv  Detail & Related papers  (2025-04-22T14:19:06Z)
• A Theoretical Framework for Data Efficient Multi-Source Transfer Learning Based on Cramér-Rao Bound [16.49737340580437]
  We propose a theoretical framework that answers the question: what is the optimal quantity of source samples needed from each source task to jointly train the target model? Specifically, we introduce a generalization error measure that aligns with cross-entropy loss, and minimize it based on the Cram'er-Rao Bound to determine the optimal transfer quantity for each source task. We develop an architecture-agnostic and data-efficient algorithm OTQMS to implement our theoretical results for training deep multi-source transfer learning models.
  arXiv  Detail & Related papers  (2025-02-06T17:32:49Z)
• SOLO: A Single Transformer for Scalable Vision-Language Modeling [74.05173379908703]
  We present SOLO, a single transformer for visiOn-Language mOdeling.<n>A unified single Transformer architecture, like SOLO, effectively addresses these scalability concerns in LVLMs.<n>In this paper, we introduce the first open-source training recipe for developing SOLO, an open-source 7B LVLM.
  arXiv  Detail & Related papers  (2024-07-08T22:40:15Z)
• PIM-Opt: Demystifying Distributed Optimization Algorithms on a Real-World Processing-In-Memory System [21.09681871279162]
  Modern Machine Learning (ML) training on large-scale datasets is a time-consuming workload. It relies on the optimization algorithm Gradient Descent (SGD) due to its effectiveness, simplicity, and generalization performance. processor-centric architectures suffer from low performance and high energy consumption while executing ML training workloads. Processing-In-Memory (PIM) is a promising solution to alleviate the data movement bottleneck.
  arXiv  Detail & Related papers  (2024-04-10T17:00:04Z)
• Diffusion-Based Neural Network Weights Generation [80.89706112736353]
  D2NWG is a diffusion-based neural network weights generation technique that efficiently produces high-performing weights for transfer learning. Our method extends generative hyper-representation learning to recast the latent diffusion paradigm for neural network weights generation. Our approach is scalable to large architectures such as large language models (LLMs), overcoming the limitations of current parameter generation techniques.
  arXiv  Detail & Related papers  (2024-02-28T08:34:23Z)
• cedar: Optimized and Unified Machine Learning Input Data Pipelines [2.0375440421573843]
  cedar is an optimized and unified programming framework for machine learning input data pipelines.<n> cedar orchestrates processing across a customizable set of local and distributed compute resources.<n> cedar improves performance by up to 1.87x to 10.65x compared to state-of-the-art input data systems.
  arXiv  Detail & Related papers  (2024-01-17T00:36:58Z)
• Bad Students Make Great Teachers: Active Learning Accelerates Large-Scale Visual Understanding [9.112203072394648]
  Power-law scaling indicates that large-scale training with uniform sampling is prohibitively slow. Active learning methods aim to increase data efficiency by prioritizing learning on the most relevant examples.
  arXiv  Detail & Related papers  (2023-12-08T19:26:13Z)
• Diffusion Model is an Effective Planner and Data Synthesizer for Multi-Task Reinforcement Learning [101.66860222415512]
  Multi-Task Diffusion Model (textscMTDiff) is a diffusion-based method that incorporates Transformer backbones and prompt learning for generative planning and data synthesis. For generative planning, we find textscMTDiff outperforms state-of-the-art algorithms across 50 tasks on Meta-World and 8 maps on Maze2D.
  arXiv  Detail & Related papers  (2023-05-29T05:20:38Z)
• Profiling and Improving the PyTorch Dataloader for high-latency Storage: A Technical Report [0.7349727826230862]
  This work focuses on the data loading pipeline in the PyTorch Framework. We show that for classification tasks that involve loading many files, like images, the training wall-time can be significantly improved. With our new, modified ConcurrentDataloader we can reach improvements in GPU utilization and significantly reduce batch loading time, up to 12X.
  arXiv  Detail & Related papers  (2022-11-09T14:16:30Z)
• An Overview of the Data-Loader Landscape: Comparative Performance Analysis [6.913175606212201]
  Dataloaders might hold the key to drastically improving the performance of training jobs. Recent advances have shown promise not only by considerably decreasing training time but also by offering new features such as loading data from remote storage like S3.
  arXiv  Detail & Related papers  (2022-09-27T21:40:56Z)
• Deep Transfer Learning for Multi-source Entity Linkage via Domain Adaptation [63.24594955429465]
  Multi-source entity linkage is critical in high-impact applications such as data cleaning and user stitching. AdaMEL is a deep transfer learning framework that learns generic high-level knowledge to perform multi-source entity linkage. Our framework achieves state-of-the-art results with 8.21% improvement on average over methods based on supervised learning.
  arXiv  Detail & Related papers  (2021-10-27T15:20:41Z)
• Multi-modal AsynDGAN: Learn From Distributed Medical Image Data without Sharing Private Information [55.866673486753115]
  We propose an extendable and elastic learning framework to preserve privacy and security. The proposed framework is named distributed Asynchronized Discriminator Generative Adrial Networks (AsynDGAN)
  arXiv  Detail & Related papers  (2020-12-15T20:41:24Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.