Model Parallelism With Subnetwork Data Parallelism

2025-07-11

AI Summary

To address high per-node memory pressure and substantial intra-node communication overhead in distributed pre-training of large-scale models, this paper proposes Subnet Data Parallel (SDP): each worker node independently trains a structured, compact subnetwork, eliminating activation transmission across pipeline stages. SDP integrates stochastic block dropping with width-wise subnetwork construction to ensure uniform parameter coverage and gradient alignment across distributed workers. Its communication bandwidth requirement is comparable to or lower than that of standard all-reduce operations. Experiments demonstrate that SDP reduces GPU memory consumption by 20–40% without compromising model accuracy, while preserving convergence properties and significantly improving training efficiency for large models.

Technology Category

Machine Learning: Distributed Machine Learning & Federated LearningSearch and Optimization: Distributed SearchData Mining & Knowledge Management: Scalability, Parallel & Distributed Systems

Application Category

Web Mining and Content Analysis: Large pretrained models with web dataGraph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphsSystems and Infrastructure for Web, Mobile and WoT: Data management and stream processing for Web, mobile and wireless applications

Abstract

Distributed pre-training of large models at scale often imposes heavy memory demands on individual nodes and incurs significant intra-node communication costs. We propose a novel alternative approach that reduces the memory requirements by training small, structured subnetworks of the model on separate workers. Unlike pipelining, our method avoids inter-node activation communication and maintains bandwidth requirements that are comparable to or lower than standard data parallel communication schemes based on all-reduce. We evaluate two subnetwork construction strategies guided by the principle of ensuring uniform representation of each parameter across the distributed training setup. Our results show that the stochastic block dropping technique consistently outperforms the width-wise subnetwork construction previously explored in federated learning. We empirically attribute this superior performance to stronger gradient alignment in subnetworks that retain blocks having skip connections. Preliminary experiments highlight the promise of our approach, achieving a 20-40% reduction in memory usage without any loss in performance.

Problem

• Reduces memory demands in distributed pre-training of large models
• Avoids inter-node activation communication with structured subnetworks
• Achieves lower memory usage without performance loss

Innovation

• Training subnetworks on separate workers
• Avoiding inter-node activation communication
• Reducing memory usage by 20-40%