StreamlineAI runs a custom PyTorch training script (script mode) that produces periodic model checkpoints. They want to use SageMaker Managed Spot Training to reduce cost. The training script writes checkpoints to /opt/ml/checkpoints and the company stores artifacts under s3://streamlineai-checkpoints/training1/. Which configuration will let a managed-spot SageMaker training job resume from the last checkpoint after an interruption?

The correct option leverages SageMaker's built-in checkpointing mechanism for Managed Spot Training. By configuring CheckpointConfig with S3Uri and LocalPath, SageMaker automatically handles the synchronization of checkpoints between the local training instance and S3. When a spot instance is interrupted, SageMaker uploads the contents of LocalPath to S3Uri. Upon restart, it downloads the latest checkpoints from S3Uri to LocalPath, allowing the training script to resume from the last saved state. The training script must be designed to save checkpoints to /opt/ml/checkpoints and load them from there upon startup. The incorrect options are: Relying on ModelArtifacts is incorrect because ModelArtifacts are typically used for the final model output at the end of a successful job, not for intermediate checkpoints to resume training. SageMaker Debugger (smdebug) is for debugging and monitoring, not for managing training checkpoints for job resumption after interruption. While it can save tensors, it's not designed for the automatic resume functionality required here. Mounting an EFS filesystem is not a standard or recommended approach for checkpointing with SageMaker Managed Spot Training. EFS is not automatically managed by SageMaker for spot interruptions in the same way CheckpointConfig is, and it introduces unnecessary complexity and potential performance overhead.