QuantGen plans to accelerate large-scale GPU training on SageMaker. They want to enable GPU kernel fusion and operator-level optimizations provided by SageMaker Training Compiler. Which statement correctly describes how to enable it and which frameworks are supported for that optimization?

The correct answer accurately describes how to enable SageMaker Training Compiler and its scope. You enable it directly on the TrainingJob or Estimator object, typically via compilerconfig={'enabled': True} or enabletrainingcompiler=True in newer SDK versions. It specifically targets GPU instances (like p3, p4, g4dn families) to perform GPU kernel fusion and graph/operator optimizations. It supports deep learning frameworks such as PyTorch and TensorFlow, which benefit from these types of optimizations. It does not apply to frameworks like scikit-learn or XGBoost, which are not deep learning frameworks and do not typically benefit from GPU kernel fusion in the same way. The incorrect options propose methods that are not valid for enabling the compiler (e.g., a hyperparameter, AutoML setting) or incorrectly state its applicability to all frameworks, CPU instances, or automatic application without configuration.