The paravirtualized device runs work on the host GPU but can report a conservative feature level. In our tested workload, selected newer paths stayed behind that reporting boundary.
Feature levels are illustrative. Apple groups GPU capability into numbered families with Metal feature sets layered on top. This graphic ends at Apple 9, the ceiling used in our test. The stock guest reported roughly Apple 5 with SIMD-group matrix support unavailable. The configured profile changed selected answers so llama.cpp could choose newer paths. Each additional feature requires its own virtualization test.