Dumb q- have you profiled the inference execution? Where are the bottlenecks you are observing?