NexaVision is evaluating distributed strategies for two workloads: (A) training ResNet50 on ImageNet across 8 x p3.8xlarge GPUs where the full model easily fits on a single GPU but they need higher effective batch size, and (B) training a 100B-parameter transformer where a single GPU cannot hold the model weights. For these two distinct workloads, choose TWO choices describing which SageMaker distributed approach is appropriate for each.

For Workload A, SMDDP is correct because the model fits on a single GPU, and data parallelism (replicating the model and distributing data) is ideal for increasing effective batch size and throughput. SMDDP leverages NCCL for efficient communication, which is crucial for scaling. The incorrect option states SMDDP cannot utilize NCCL for image models, which is false; SMDDP is built on torch.distributed and uses NCCL. For Workload B, Model Parallel (SageMaker MP/pipeline parallel) is correct because the 100B-parameter model is too large for a single GPU. Model parallelism shards the model parameters across multiple devices, allowing training of models that exceed single-device memory. The incorrect option suggests SMDDP for Workload B, which is unsuitable as SMDDP replicates the entire model, and a 100B-parameter model would not fit on each GPU. Single-node training is incorrect for both as it doesn't address the distributed needs.