You are packaging a custom ResNet training environment into a Docker container to run on GPU instances (NVIDIA) in SageMaker (using EC2 P3 instances). What must you do so the container can use the GPUs properly?

To properly utilize GPUs within a Docker container on NVIDIA hardware, the container must be built to be compatible with the NVIDIA Container Toolkit (formerly NVIDIA-Docker). This toolkit provides the necessary runtime and libraries to allow Docker containers to access the host's NVIDIA GPUs and their drivers. SageMaker leverages this capability when running training jobs on GPU instances. Including NVIDIA drivers inside the Docker image is incorrect because the drivers are typically installed on the host operating system, not within the container itself. The NVIDIA Container Toolkit bridges the container to these host drivers. Arranging the container file layout specifically for GPU instances is too vague and doesn't address the core technical requirement of GPU access. Setting a GPU flag in the SageMaker CreateTrainingJob API request is necessary to tell SageMaker to provision a GPU instance, but it doesn't configure the container itself for GPU access.