medium10 pts · first answer countsgpu-computingprofiling

Finding the slow part of a step

A training step takes 900 ms and you do not know why. What is the best first tool to see time per operator, on CPU and GPU?

A.Adding print statements to the modelB.Doubling the batch size and timing againC.Reading the nvidia-smi output onceD.torch.profiler with a trace viewer

Discussion 0

Join free to post in the discussion.

Finding the slow part of a step

Discussion 0

Related questions