Are the times reported by tvm.runtime.profiler_vm.VirtualMachineProfiler's profile() function the actual times observed on running inference on the target device?
I’m not really sure what you mean by “actual times”. The per op times reported by the profiler are the time it takes to run the op along with the time it takes to start and stop the timer. The overall times ate the end are the time from the start of the model execution to the end, including the time it takes to run profilers for all the ops and collect the shape information. Note that these timers are device specific. i.e. the GPU timer only times how long the op takes to execute on the GPU and not the time it takes to synchronize with the host.