I am learning about using the
torch.profiler.schedule API. But from its introduction, I cannot figure out the specific difference between wait steps and warmup steps inside it.
What I can get it that the profiler would not do recording in both wait steps and warmup steps, but is there any other specific differences of operations between them?
In other words, why are the wait steps are esstianl for profiling?
Your answer and guide will be appreciated!