llm-helper metal-limit

Show or raise how much unified memory macOS lets the GPU use

macOS caps how much unified memory the GPU may wire, and vllm-metal sizes --gpu-memory-utilization as a share of that cap, not of installed RAM. On a Mac dedicated to serving models the default leaves memory idle.

With no flags, show the current cap and a suggested one. --set raises it now and installs a LaunchDaemon that applies it again at every boot, since macOS resets it on restart. The server reads the cap once at startup, so run 'llm-helper restart' afterwards. macOS only.

Usage

llm-helper metal-limit [flags]

Options

--reset

return to the macOS default and remove the boot-time daemon

--set

set the cap to this many MiB, or 'auto' for the suggested value

--yes

apply without asking