The team is running consecutive training jobs in SageMaker and wants to minimize infrastructure startup time for each job. Which option will reduce startup latency?

SageMaker managed warm pools are designed to reduce infrastructure startup time by keeping a pool of pre-initialized training instances ready for use. When a new training job starts, SageMaker can quickly allocate an instance from this warm pool, significantly cutting down the time it takes to provision and set up the environment. Managed Spot Training focuses on cost reduction by using spare EC2 capacity, but it does not inherently reduce startup time; in fact, spot instances can sometimes have longer startup times due to provisioning. The SageMaker Training Compiler optimizes the training graph for faster execution after the instance has started, not during startup. The SageMaker distributed data parallelism (SMDDP) library accelerates the training process itself by distributing the workload across multiple GPUs or instances, but it does not address the initial infrastructure startup latency.