Description
When using on-demand profiling via dynolog and kineto, we noticed that, when profiling request configured with iterations, the last profiling iteration took more time than other profiling iterations. The train process is blocked at optimizer.step(), which calls step in kineto, finally, in performRunLoop, libkineto::api().client()->stop() took the most time.
At the same time, the processTraceInternal is executed asynchronously in performRunLoop, which will not block torch train process.
I'm wondering whether there's a plan to fix this performance issue to make minimal overhead on pytorch training process when on-demand profiling is enabled. it would be very nice if there's already a plan or a proposal. If not, I'd like to make a proposal later.