Optimizing Distributed Training on Frontier for Large Language Models

Abstract: Large language models (LLM) are showing tremendous success as foundation models, and many downstream applications benefit from fine-tuning. Prior works on loss scaling have demonstrated that the larger LLMs perform better than their smaller counterparts. However, training LLMs with billions of parameters requires considerable computational resources; to train a one trillion GPT-style model on 20 trillion tokens, we need to perform 120 million exaflops. Frontier is the world's first and fastest exascale supercomputer for open science and is equipped with 75264 MI250X GPUs. This work explores efficient distributed strategies such as tensor parallelism, pipeline parallelism, and sharded data parallelism to train a trillion-parameter model on the Frontier exascale supercomputer. We analyze these distributed training techniques and associated parameters individually to decide which techniques to use and what associated parameters to select for a particular technique. We perform hyperparameter tuning on these techniques to understand their complex interplay. Combined with these two tuning efforts, we have found optimal strategies to train three models of size 22B, 175B, and 1T parameters with $38.38\%$ , $36.14\%$ , and $31.96\%$ achieved throughput. For training the 175B parameter model and 1T model, we have achieved $100\%$ weak scaling efficiency and $89\%$ and $87\%$ strong scaling efficiency, respectively. Our work presents a set of strategies for distributed training of LLMs through experimental findings and hyperparameter tuning.