A county assessor must train a large imagery model that no longer fits practical single-node GPU limits. What should the team use in Azure Machine Learning?

Managing distributed training across multiple GPUs for the large deep-learning imagery model is the Azure Machine Learning approach when county assessor workloads exceed practical single-node GPU limits. Distributed training frameworks and multi-node/multi-GPU compute clusters partition work so model size and throughput remain feasible. Domain 2 covers distributed training as the response to large and deep learning jobs rather than enabling MLflow autologging with distributed run tracking for the large deep-learning imagery model, which improves observability but does not add the GPU capacity needed for deep-model training.

Running Hyperdrive sweep jobs across single-GPU nodes for the large deep-learning imagery model fails by concept: Hyperdrive launches separate hyperparameter trials that each remain constrained by single-node GPU limits and does not partition and jointly train one oversized model across multiple GPUs. Enabling MLflow autologging with distributed run tracking for the large deep-learning imagery model fails methodology: autologging only records metrics, parameters, and artifacts for observability and does not supply the distributed GPU compute required when a model exceeds single-node limits. Running Automated ML vision tasks on a CPU compute cluster for the large deep-learning imagery model fails fit and capacity: Automated ML on a CPU cluster does not provide multi-GPU distributed training and cannot accommodate a large deep-learning model that no longer fits practical single-node GPU limits.

Exam caveat: Large/deep learning capacity questions point to distributed training in Azure ML; running Hyperdrive sweeps on single-GPU nodes, enabling MLflow autologging with run tracking, and running Automated ML vision tasks on a CPU cluster are distractors.

Operational check: configure a multi-GPU compute cluster, launch a distributed training job with the imagery script, confirm workers start and gradients synchronize as expected, monitor GPU utilization, and compare time-to-epoch against the prior single-node baseline.