Metal Execution Path

From conservative capability answers to faster Metal kernels

The host GPU and Apple’s virtualization bridge stay the same. The decision changes when one guest process asks which Metal paths it can use.

Host Apple GPU

Executes the workload on Apple Silicon

Unchanged

Virtualization.framework bridge

Apple’s existing paravirtualized graphics path

Unchanged

Guest paravirtualized GPU

The same virtual Metal device in both runs

Unchanged

Capability query

What can this workload safely select?

One process

Stock guest answer

Apple 5 · 32 KB

SIMD-group matrix unavailable
Conservative feature selection

→slower llama.cpp kernel

Shim · tested profile

Apple 9 · 64 KB

SIMD-group matrix · bfloat16
Newer tested Metal paths

→faster llama.cpp kernel

Selected capability answers change for one guest process while Apple’s bridge and virtual device stay in place. Each additional Metal API requires its own virtualization test.