A team is training a large language model using SageMaker with multiple GPUs. They need to reduce training time by splitting the model across devices due to memory constraints. Which distributed training strategy should they use?