Data Parallelism Quiz

Answers and explanations appear as you take the quiz.

1. In standard data parallelism, what does each GPU hold?
   • A complete copy of the model, processing a different slice of the batch
   • A different slice of the model's layers
   • Only the optimizer state, with weights stored centrally
   • A compressed version of the model with fewer parameters
2. Which communication operation keeps the model replicas in sync each step?
   • Broadcast, which copies one GPU's data to all others
   • All-reduce, which averages gradients across all GPUs
   • Scatter, which divides data among GPUs
   • No communication is needed between GPUs
3. What memory limit remains in plain data parallelism?
   • The full model is sharded automatically across the GPUs
   • Only the input batch uses device memory
   • Each GPU keeps model parameters and optimizer state replicated, limiting the model size that fits
   • Each worker stores different model layers by default
4. Which factors does NCCL use when selecting collective algorithms?
   • Only the random seed and learning rate
   • Only the total number of model parameters
   • The optimizer type and validation score
   • Message size and topology
5. How does PyTorch DistributedDataParallel hide communication latency?
   • By overlapping gradient all-reduce with the ongoing backward pass
   • By skipping gradient synchronization on alternating steps
   • By compressing all gradients to a single number
   • By running the forward pass on the CPU
6. With equal local batch size B on each of W data-parallel workers, what is the global batch for one synchronized step?
   • B divided by W
   • B multiplied by W, before any gradient-accumulation factor
   • Always exactly W samples
   • The model parameter count divided by B
7. Why may adding more data-parallel GPUs fail to increase throughput proportionally?
   • All-reduce makes model parameters disappear
   • The model is copied only to one worker
   • Gradient communication, data loading, and synchronization add overhead
   • More workers always reduce the global batch to zero
8. What does gradient accumulation let a data-parallel training loop do?
   • Shard parameters automatically across all workers
   • Skip forward and backward computation
   • Replace the model with an all-reduce operation
   • Combine gradients from multiple microbatches before an optimizer update