Optimizing Distributed Training on Frontier for Large Language Models