Yeah that layer looping is super interesting. Also, it’s memory friendly with the right inference harness.