Question 190
A distributed training run writes a separate checkpoint from every rank and corrupts the shared target. The next test must distinguish a real fix from a configuration change that leaves the symptom intact. A release test must show that exactly one valid checkpoint artifact appears for the training step. Which next step fixes the reported cause?
Explanation
Correct Answer
Concurrent workers can overwrite one checkpoint location without coordinated file writes. The expected outcome is that exactly one valid checkpoint artifact appears for the training step.
Incorrect Answers
Data-parallel loaders must divide samples so workers do not repeatedly process identical mini-batches.
Distributed training coordinates work across multiple GPU processes rather than adding CPU threads alone.
Collective operations require reachable peer communication and coordinated initialization.