Asynchronous Stochastic Gradient Descent with Decoupled Backpropagation and Layer-Wise Updates

• URL: http://arxiv.org/abs/2410.05985v3
• Date: Fri, 07 Feb 2025 13:33:12 GMT
• Title: Asynchronous Stochastic Gradient Descent with Decoupled Backpropagation and Layer-Wise Updates
• Authors: Cabrel Teguemne Fokam, Khaleelulla Khan Nazeer, Lukas König, David Kappel, Anand Subramoney,
• Abstract summary: Asynchronous gradient descent (ASGD) methods can improve training speed, but are sensitive to delays due to both communication and throughput differences. PD-ASGD uses separate threads for the forward and backward passes, decoupling the updates and allowing for a higher ratio of forward to backward threads. Our approach yields close to state-of-the-art results while running up to $5.95times$ faster than synchronous data parallelism in the presence of delays.
• Score: 1.9241821314180372
• License:
• Abstract: The increasing size of deep learning models has made distributed training across multiple devices essential. However, current methods such as distributed data-parallel training suffer from large communication and synchronization overheads when training across devices, leading to longer training times as a result of suboptimal hardware utilization. Asynchronous stochastic gradient descent (ASGD) methods can improve training speed, but are sensitive to delays due to both communication and differences throughput. Moreover, the backpropagation algorithm used within ASGD workers is bottlenecked by the interlocking between its forward and backward passes. Current methods also do not take advantage of the large differences in the computation required for the forward and backward passes. Therefore, we propose an extension to ASGD called Partial Decoupled ASGD (PD-ASGD) that addresses these issues. PD-ASGD uses separate threads for the forward and backward passes, decoupling the updates and allowing for a higher ratio of forward to backward threads than the usual 1:1 ratio, leading to higher throughput. PD-ASGD also performs layer-wise (partial) model updates concurrently across multiple threads. This reduces parameter staleness and consequently improves robustness to delays. Our approach yields close to state-of-the-art results while running up to $5.95\times$ faster than synchronous data parallelism in the presence of delays, and up to $2.14\times$ times faster than comparable ASGD algorithms by achieving higher model flops utilization. We mathematically describe the gradient bias introduced by our method, establish an upper bound, and prove convergence.

Related papers

• BitPipe: Bidirectional Interleaved Pipeline Parallelism for Accelerating Large Models Training [5.7294516069851475]
  BitPipe is a bidirectional interleaved pipeline parallelism for accelerating large models training. We show that BitPipe improves the training throughput of GPT-style and BERT-style models by 1.05x-1.28x compared to the state-of-the-art synchronous approaches.
  arXiv  Detail & Related papers  (2024-10-25T08:08:51Z)
• AsyncDiff: Parallelizing Diffusion Models by Asynchronous Denoising [49.785626309848276]
  AsyncDiff is a universal and plug-and-play acceleration scheme that enables model parallelism across multiple devices. For the Stable Diffusion v2.1, AsyncDiff achieves a 2.7x speedup with negligible degradation and a 4.0x speedup with only a slight reduction of 0.38 in CLIP Score. Our experiments also demonstrate that AsyncDiff can be readily applied to video diffusion models with encouraging performances.
  arXiv  Detail & Related papers  (2024-06-11T03:09:37Z)
• Inverse-Free Fast Natural Gradient Descent Method for Deep Learning [52.0693420699086]
  We present a fast natural gradient descent (FNGD) method that only requires inversion during the first epoch. FNGD exhibits similarities to the average sum in first-order methods, leading to the computational complexity of FNGD being comparable to that of first-order methods.
  arXiv  Detail & Related papers  (2024-03-06T05:13:28Z)
• Robust Fully-Asynchronous Methods for Distributed Training over General Architecture [11.480605289411807]
  Perfect synchronization in distributed machine learning problems is inefficient and even impossible due to the existence of latency, package losses and stragglers. We propose Fully-Asynchronous Gradient Tracking method (R-FAST), where each device performs local computation and communication at its own without any form of impact.
  arXiv  Detail & Related papers  (2023-07-21T14:36:40Z)
• Delay-adaptive step-sizes for asynchronous learning [8.272788656521415]
  We show that it is possible to use learning rates that depend on the actual time-varying delays in the system. For each of these methods, we demonstrate how delays can be measured on-line, present delay-adaptive step-size policies, and illustrate their theoretical and practical advantages over the state-of-the-art.
  arXiv  Detail & Related papers  (2022-02-17T09:51:22Z)
• Distributed stochastic optimization with large delays [59.95552973784946]
  One of the most widely used methods for solving large-scale optimization problems is distributed asynchronous gradient descent (DASGD) We show that DASGD converges to a global optimal implementation model under same delay assumptions.
  arXiv  Detail & Related papers  (2021-07-06T21:59:49Z)
• Gradient Coding with Dynamic Clustering for Straggler Mitigation [57.9123881133818]
  GC-DC regulates the number of straggling workers in each cluster based on the straggler behavior in the previous iteration. We numerically show that GC-DC provides significant improvements in the average completion time (of each iteration) with no increase in the communication load compared to the original GC scheme.
  arXiv  Detail & Related papers  (2020-11-03T18:52:15Z)
• Stochastic Optimization with Laggard Data Pipelines [65.20044914532221]
  We show that "dataechoed" extensions of common optimization methods exhibit provable improvements over their synchronous counterparts. Specifically, we show that in convex optimization with minibatches, data echoing affords speedups on the curvature-dominated part of the convergence rate, while maintaining the optimal statistical rate.
  arXiv  Detail & Related papers  (2020-10-26T14:55:31Z)
• Sparse Communication for Training Deep Networks [56.441077560085475]
  Synchronous gradient descent (SGD) is the most common method used for distributed training of deep learning models. In this algorithm, each worker shares its local gradients with others and updates the parameters using the average gradients of all workers. We study several compression schemes and identify how three key parameters affect the performance.
  arXiv  Detail & Related papers  (2020-09-19T17:28:11Z)
• 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)
• OD-SGD: One-step Delay Stochastic Gradient Descent for Distributed Training [5.888925582071453]
  We propose a novel technology named One-step Delay SGD (OD-SGD) to combine their strengths in the training process. We evaluate our proposed algorithm on MNIST, CIFAR-10 and ImageNet datasets.
  arXiv  Detail & Related papers  (2020-05-14T05:33:36Z)

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.