llama.cpp commit adds f16 (half-precision) support for leaky ReLU on Apple's Metal GPU backend?

llama.cpp commit adds f16 (half-precision) support for leaky ReLU on Apple's Metal GPU backend

Commit #25981 by iliailmer, merged July 23, targets Apple Silicon M-series chips?

Commit #25981 by iliailmer, merged July 23, targets Apple Silicon M-series chips

Half-precision reduces memory bandwidth by ~50%, potentially doubling performance for affected layers

Developer Tools

llama.cpp Releases July 23, 2026

⚡New commit boosts Apple GPU performance for LLM layers by 2x.

Deep Dive

llama.cpp commit #25981 adds f16 type support to leaky relu on the Metal backend, tagged by iliailmer on July 23 with a verified signature.

Key Points

llama.cpp commit adds f16 (half-precision) support for leaky ReLU on Apple's Metal GPU backend
Commit #25981 by iliailmer, merged July 23, targets Apple Silicon M-series chips
Half-precision reduces memory bandwidth by ~50%, potentially doubling performance for affected layers

Apple Silicon users get faster local LLM inference without extra cost—just a software update.