Asynchronous Local-SGD Training for Language Modeling

• URL: http://arxiv.org/abs/2401.09135v1
• Date: Wed, 17 Jan 2024 11:17:04 GMT
• Title: Asynchronous Local-SGD Training for Language Modeling
• Authors: Bo Liu, Rachita Chhaparia, Arthur Douillard, Satyen Kale, Andrei A. Rusu, Jiajun Shen, Arthur Szlam, Marc'Aurelio Ranzato
• Abstract summary: Local gradient descent (Local-SGD) is an approach to distributed optimization where each device performs more than one SGD update per communication. This work presents an empirical study of it asynchronous Local-SGD for training language models; that is, each worker updates the global parameters as soon as it has finished its SGD steps.
• Score: 38.61892210645179
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Local stochastic gradient descent (Local-SGD), also referred to as federated averaging, is an approach to distributed optimization where each device performs more than one SGD update per communication. This work presents an empirical study of {\it asynchronous} Local-SGD for training language models; that is, each worker updates the global parameters as soon as it has finished its SGD steps. We conduct a comprehensive investigation by examining how worker hardware heterogeneity, model size, number of workers, and optimizer could impact the learning performance. We find that with naive implementations, asynchronous Local-SGD takes more iterations to converge than its synchronous counterpart despite updating the (global) model parameters more frequently. We identify momentum acceleration on the global parameters when worker gradients are stale as a key challenge. We propose a novel method that utilizes a delayed Nesterov momentum update and adjusts the workers' local training steps based on their computation speed. This approach, evaluated with models up to 150M parameters on the C4 dataset, matches the performance of synchronous Local-SGD in terms of perplexity per update step, and significantly surpasses it in terms of wall clock time.

Related papers

• PeFAD: A Parameter-Efficient Federated Framework for Time Series Anomaly Detection [51.20479454379662]
  We propose a. Federated Anomaly Detection framework named PeFAD with the increasing privacy concerns. We conduct extensive evaluations on four real datasets, where PeFAD outperforms existing state-of-the-art baselines by up to 28.74%.
  arXiv  Detail & Related papers  (2024-06-04T13:51:08Z)
• Stragglers-Aware Low-Latency Synchronous Federated Learning via Layer-Wise Model Updates [71.81037644563217]
  Synchronous federated learning (FL) is a popular paradigm for collaborative edge learning. As some of the devices may have limited computational resources and varying availability, FL latency is highly sensitive to stragglers. We propose straggler-aware layer-wise federated learning (SALF) that leverages the optimization procedure of NNs via backpropagation to update the global model in a layer-wise fashion.
  arXiv  Detail & Related papers  (2024-03-27T09:14:36Z)
• Mitigating System Bias in Resource Constrained Asynchronous Federated Learning Systems [2.8790600498444032]
  We propose a dynamic global model aggregation method within Asynchronous Federated Learning (AFL) deployments. Our method scores and adjusts the weighting of client model updates based on their upload frequency to accommodate differences in device capabilities.
  arXiv  Detail & Related papers  (2024-01-24T10:51:15Z)
• Federated Learning with Projected Trajectory Regularization [65.6266768678291]
  Federated learning enables joint training of machine learning models from distributed clients without sharing their local data. One key challenge in federated learning is to handle non-identically distributed data across the clients. We propose a novel federated learning framework with projected trajectory regularization (FedPTR) for tackling the data issue.
  arXiv  Detail & Related papers  (2023-12-22T02:12:08Z)
• Straggler-Resilient Decentralized Learning via Adaptive Asynchronous Updates [28.813671194939225]
  fully decentralized optimization methods have been advocated as alternatives to the popular parameter server framework. We propose a fully decentralized algorithm with adaptive asynchronous updates via adaptively determining the number of neighbor workers for each worker to communicate with. We show that DSGD-AAU achieves a linear speedup for convergence and demonstrate its effectiveness via extensive experiments.
  arXiv  Detail & Related papers  (2023-06-11T02:08:59Z)
• STSyn: Speeding Up Local SGD with Straggler-Tolerant Synchronization [14.526055067546507]
  Local synchronization suffers from some workers being idle random delays due to slow and straggler workers, as it waits for the workers to complete the same amount of local updates. In this paper, to mitigate stragglers and improve communication efficiency, a novel local SGD system strategy, named STSyn, is developed.
  arXiv  Detail & Related papers  (2022-10-06T08:04:20Z)
• Gradient Coding with Dynamic Clustering for Straggler-Tolerant Distributed Learning [55.052517095437]
  gradient descent (GD) is widely employed to parallelize the learning task by distributing the dataset across multiple workers. A significant performance bottleneck for the per-iteration completion time in distributed synchronous GD is $straggling$ workers. Coded distributed techniques have been introduced recently to mitigate stragglers and to speed up GD iterations by assigning redundant computations to workers. We propose a novel dynamic GC scheme, which assigns redundant data to workers to acquire the flexibility to choose from among a set of possible codes depending on the past straggling behavior.
  arXiv  Detail & Related papers  (2021-03-01T18:51:29Z)
• HPSGD: Hierarchical Parallel SGD With Stale Gradients Featuring [18.8426865970643]
  A novel Hierarchical Parallel SGD (HPSGD) strategy is proposed to boost the distributed training process of the deep neural network (DNN) Experiments are conducted to demonstrate that the proposed HPSGD approach substantially boosts the distributed DNN training, reduces the disturbance of the stale gradients and achieves better accuracy in given fixed wall-time.
  arXiv  Detail & Related papers  (2020-09-06T10:17:56Z)
• DaSGD: Squeezing SGD Parallelization Performance in Distributed Training Using Delayed Averaging [4.652668321425679]
  Minibatch gradient descent (SGD) algorithm requires workers to halt forward/back propagations. DaSGD parallelizes SGD and forward/back propagations to hide 100% of the communication overhead.
  arXiv  Detail & Related papers  (2020-05-31T05:43:50Z)
• Variance Reduced Local SGD with Lower Communication Complexity [52.44473777232414]
  We propose Variance Reduced Local SGD to further reduce the communication complexity. VRL-SGD achieves a emphlinear iteration speedup with a lower communication complexity $O(Tfrac12 Nfrac32)$ even if workers access non-identical datasets.
  arXiv  Detail & Related papers  (2019-12-30T08:15:21Z)

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.