You have implemented a custom training algorithm inside a Docker container that Amazon SageMaker will use. How should you package the container so SageMaker can start training correctly?

SageMaker expects the Docker container's ENTRYPOINT to execute the training program. When SageMaker launches the container, it will run the command specified in the ENTRYPOINT to initiate the training job. This allows SageMaker to manage the lifecycle of your training process. Adding a server argument to ENTRYPOINT is incorrect because training containers do not typically run a server; that's for inference containers. Including the path to training data in the docker build command is incorrect because training data is mounted into the container at runtime by SageMaker, not built into the image. Using a COPY instruction to place the training program under /opt/ml/train is incorrect because while the training script should be copied into the container, its specific location isn't what tells SageMaker to execute it; the ENTRYPOINT instruction does that.