llama.cpp turns on the Metal 4.0 tensor API for M5 and A19 chips, loading those kernels from a separate metallib

GitHub·medium signalXBlueskyLinkedInCopy link

Build b10734, published 2026-09-01 (#27461), requests the Metal 4.0 language version to reach Apple's tensor API on M5+ and A19+ silicon, and loads the tensor API kernels from a separate metallib rather than the main one so older devices are unaffected. The change ships with an external-metallib tensor API regression test and a fix to the metallib build order. This is the first llama.cpp build to target Apple's newest matrix hardware path, which matters for anyone running local inference on current-generation Macs and iPhones.