The Parthenon framework, derived from Athena++ (ascl:1912.005), handles massively-parallel, device-accelerated adaptive mesh refinement. It provides a device first/device resident approach, transparent packing of data across blocks (to reduce/hide kernel launch latency), and direct device-to-device communication via asynchronous, one-sided MPI communication to enable high performance. Parthenon uses an intermediate abstraction layer to hide complexity of device kernel launches, offers support for particles and abstract variable control via metadata tags, and has a flexible plug-in package system.