There’s a “profiling graph executor” that records shapes and then hands them off to a fusion compiler. The profiling executor re-specializes on every new shape

There’s a “profiling graph executor” that records shapes and then hands them off to a fusion compiler. The profiling executor re-specializes on every new shape it sees, but stops at 20 re-specialization. We’re working on eliminating the dependence on shape specialization right now, since it’s kind of an unfortunate limitation for some workloads.