Question 250
A large PyTorch model repeatedly overflows gradients when training with half-precision arithmetic, producing NaN losses, while a full-precision baseline converges. Which training control should be evaluated?
Explanation
Correct Answer
Reduced precision has a narrower numeric range; loss scaling and compatible mixed-precision tooling can mitigate underflow or overflow when supported.
Incorrect Answers
Endpoint access is unrelated to training numerical stability.
Registry retention has no effect on gradient arithmetic.
Suppressing logs conceals the failure.