ACCO: Accumulate while you Communicate, Hiding Communications in Distributed LLM Training

• URL: http://arxiv.org/abs/2406.02613v1
• Date: Mon, 3 Jun 2024 08:23:45 GMT
• Title: ACCO: Accumulate while you Communicate, Hiding Communications in Distributed LLM Training
• Authors: Adel Nabli, Louis Fournier, Pierre Erbacher, Louis Serrano, Eugene Belilovsky, Edouard Oyallon,
• Abstract summary: We propose a memory-efficient optimization algorithm tailored for distributed training of Large Language Models. Our method relies on a novel technique to mitigate the one-step delay inherent in parallel execution of gradient computations and communications.
• Score: 16.560270624096706
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Training Large Language Models (LLMs) relies heavily on distributed implementations, employing multiple GPUs to compute stochastic gradients on model replicas in parallel. However, synchronizing gradients in data parallel settings induces a communication overhead increasing with the number of distributed workers, which can impede the efficiency gains of parallelization. To address this challenge, optimization algorithms reducing inter-worker communication have emerged, such as local optimization methods used in Federated Learning. While effective in minimizing communication overhead, these methods incur significant memory costs, hindering scalability: in addition to extra momentum variables, if communications are only allowed between multiple local optimization steps, then the optimizer's states cannot be sharded among workers. In response, we propose $\textbf{AC}$cumulate while $\textbf{CO}$mmunicate ($\texttt{ACCO}$), a memory-efficient optimization algorithm tailored for distributed training of LLMs. $\texttt{ACCO}$ allows to shard optimizer states across workers, overlaps gradient computations and communications to conceal communication costs, and accommodates heterogeneous hardware. Our method relies on a novel technique to mitigate the one-step delay inherent in parallel execution of gradient computations and communications, eliminating the need for warmup steps and aligning with the training dynamics of standard distributed optimization while converging faster in terms of wall-clock time. We demonstrate the effectiveness of $\texttt{ACCO}$ on several LLMs training and fine-tuning tasks.

Related papers

• A Quadratic Synchronization Rule for Distributed Deep Learning [66.68264684667562]
  This work proposes a theory-grounded method for determining $H$, named the Quadratic Synchronization Rule (QSR) Experiments on ResNet and ViT show that local gradient methods with QSR consistently improve the test accuracy over other synchronization strategies.
  arXiv  Detail & Related papers  (2023-10-22T21:38:57Z)
• Accelerating Distributed ML Training via Selective Synchronization [0.0]
  textttSelSync is a practical, low-overhead method for DNN training that dynamically chooses to incur or avoid communication at each step. Our system converges to the same or better accuracy than BSP while reducing training time by up to 14$times$.
  arXiv  Detail & Related papers  (2023-07-16T05:28:59Z)
• $\textbf{A}^2\textbf{CiD}^2$: Accelerating Asynchronous Communication in Decentralized Deep Learning [0.0]
  We introduce a principled asynchronous, randomized, gossip-based optimization algorithm which works thanks to a continuous local momentum named $textbfA2textbfCiD2$. Our theoretical analysis proves accelerated rates compared to previous asynchronous decentralized baselines. We show consistent improvement on the ImageNet dataset using up to 64 asynchronous workers.
  arXiv  Detail & Related papers  (2023-06-14T06:52:07Z)
• TAMUNA: Doubly Accelerated Distributed Optimization with Local Training, Compression, and Partial Participation [53.84175614198885]
  In distributed optimization and learning, several machines alternate between local computations in parallel and communication with a distant server. We propose TAMUNA, the first algorithm for distributed optimization that leveraged the two strategies of local training and compression jointly and allows for partial participation.
  arXiv  Detail & Related papers  (2023-02-20T08:37:44Z)
• Communication-Efficient Adam-Type Algorithms for Distributed Data Mining [93.50424502011626]
  We propose a class of novel distributed Adam-type algorithms (emphi.e., SketchedAMSGrad) utilizing sketching. Our new algorithm achieves a fast convergence rate of $O(frac1sqrtnT + frac1(k/d)2 T)$ with the communication cost of $O(k log(d))$ at each iteration.
  arXiv  Detail & Related papers  (2022-10-14T01:42:05Z)
• DADAO: Decoupled Accelerated Decentralized Asynchronous Optimization [0.0]
  DADAO is the first decentralized, accelerated, asynchronous, primal, first-order algorithm to minimize a sum of $L$-smooth and $mu$-strongly convex functions distributed over a given network of size $n$. We show that our algorithm requires $mathcalO(nsqrtchisqrtfracLmulog(frac1epsilon)$ local and only $mathcalO(nsqrtchisqrtfracLmulog(
  arXiv  Detail & Related papers  (2022-07-26T08:47:54Z)
• AsySQN: Faster Vertical Federated Learning Algorithms with Better Computation Resource Utilization [159.75564904944707]
  We propose an asynchronous quasi-Newton (AsySQN) framework for vertical federated learning (VFL) The proposed algorithms make descent steps scaled by approximate without calculating the inverse Hessian matrix explicitly. We show that the adopted asynchronous computation can make better use of the computation resource.
  arXiv  Detail & Related papers  (2021-09-26T07:56:10Z)
• Sparse Communication for Training Deep Networks [56.441077560085475]
  Synchronous gradient descent (SGD) is the most common method used for distributed training of deep learning models. In this algorithm, each worker shares its local gradients with others and updates the parameters using the average gradients of all workers. We study several compression schemes and identify how three key parameters affect the performance.
  arXiv  Detail & Related papers  (2020-09-19T17:28:11Z)
• Straggler-aware Distributed Learning: Communication Computation Latency Trade-off [56.08535873173518]
  Straggling workers can be tolerated by assigning redundant computations and coding across data and computations. In most existing schemes, each non-straggling worker transmits one message per iteration to the parameter server (PS) after completing all its computations. Imposing such a limitation results in two main drawbacks; over-computation due to inaccurate prediction of the straggling behaviour, and under-utilization due to treating workers as straggler/non-straggler.
  arXiv  Detail & Related papers  (2020-04-10T08:39:36Z)
• Communication Contention Aware Scheduling of Multiple Deep Learning Training Jobs [17.45154289084637]
  We establish a new DDL job scheduling framework which organizes DDL jobs as Directed Acyclic Graphs (DAGs) We then propose an efficient algorithm, LWF-$kappa$, to balance the GPU utilization and consolidate the allocated GPU for each job. We show that LWF-$kappa$ achieves up to $1.59times$ improvement over the classical first-fit algorithms.
  arXiv  Detail & Related papers  (2020-02-24T07:50:56Z)

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.