Galvatron: Efficient Transformer Training over Multiple GPUs Using Automatic Parallelism

Abstract

Transformer models have achieved state-of-the-art performance on various domains of applications and gradually becomes the foundations of the advanced large deep learning (DL) models. However, how to train these models over multiple GPUs efficiently is still challenging due to a large number of parallelism choices. Existing DL systems either rely on manual efforts to make distributed training plans or apply parallelism combinations within a very limited search space. In this approach, we propose Galvatron, a new system framework that incorporates multiple popular parallelism dimensions and automatically finds the most efficient hybrid parallelism strategy. To better explore such a rarely huge search space, we 1) involve a decision tree to make decomposition and pruning based on some reasonable intuitions, and then 2) design a dynamic programming search algorithm to generate the optimal plan. Evaluations on four representative Transformer workloads show that Galvatron could perform automatically distributed training with different GPU memory budgets. Among all evluated scenarios, Galvatron always achieves superior system throughput compared to previous work with limited parallelism.

References in corpus (6)

• Supporting Very Large Models using Automatic Dataflow Graph Partitioning
• Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning
• HET: Scaling out Huge Embedding Model Training via Cache-enabled Distributed Framework
• TensorOpt: Exploring the Tradeoffs in Distributed DNN Training with Auto-Parallelism
• Enabling Compute-Communication Overlap in Distributed Deep Learning Training Platforms
• Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM

Cited by in corpus (10)

• Understanding the Potential of FPGA-Based Spatial Acceleration for Large Language Model Inference
• FlexMoE: Scaling Large-scale Sparse Pre-trained Model Training via Dynamic Device Placement
• Improving Automatic Parallel Training via Balanced Memory Workload Optimization
• Mist: Efficient Distributed Training of Large Language Models via Memory-Parallelism Co-Optimization
• Apt-Serve: Adaptive Request Scheduling on Hybrid Cache for Scalable LLM Inference Serving
• Hydraulis: Balancing Large Transformer Model Training via Co-designing Parallel Strategies and Data Assignment
• LobRA: Multi-tenant Fine-tuning over Heterogeneous Data
• MEMO: Fine-grained Tensor Management For Ultra-long Context LLM Training
• Oases: Efficient Large-Scale Model Training on Commodity Servers via Overlapped and Automated Tensor Model Parallelism
• DFLOP: A Data-driven Framework for Multimodal LLM Training Pipeline Optimization