A research lab must train a multi-billion-parameter language model across hundreds of accelerator instances for several weeks. Hardware failures must be detected and faulty nodes replaced automatically without restarting the whole run from scratch. Which SageMaker option fits?
Show answer and explanation
Answer: D
SageMaker HyperPod provisions persistent GPU clusters for large distributed training, continuously checks node health, and replaces faulty instances automatically while preserving checkpoints so training can resume. Ordinary training jobs are ephemeral and suited to shorter runs. Processing jobs handle data preparation and evaluation, Batch Transform performs offline inference, and notebook instances are single interactive machines not meant for multi-week distributed training.
Why the other options are wrong
- A. Processing jobs run data transformation or evaluation scripts on a managed cluster that is released afterward. They are not designed for multi-week distributed training with automatic replacement of failed accelerators.
- B. Batch Transform runs offline inference over a dataset using a trained model. It does not train models and has no node health management for long training campaigns.
- C. Notebook instances are interactive single-instance environments used for exploration and small experiments. They cannot coordinate hundreds of accelerators or recover a distributed run after a node fails.