Mac Metal GPU Access Fails Over SSH, Fix Requires GUI Domain Relaunch

arthurcolle · x · 2026-08-16

Running llama-server via plain SSH on a Mac resulted in slow inference (5 tok/s) despite the Metal dylib being linked.

Root Cause:

When a process is launched over SSH without a GUI session context, Metal cannot detect a GPU device, forcing a fallback to CPU-only mode (logs showed nthreads=8 and zero Metal init). LM Studio works because it runs inside the logged-in GUI session.

Fix:

Relaunch the llama-server in the GUI domain using launchctl to ensure it inherits the necessary Metal context and GPU acceleration.

Original post →

More from Infra

Infra channel →