llama.cpp removes --mmap, --mlock and --direct-io; use --load-mode

▼ WHAT HAPPENED

llama.cpp merged PR #28334 on September 9, 2026, removing `--mmap`, `--no-mmap`, `--mlock`, `--direct-io` and their negative forms from the argument parser. The flags had been deprecated since PR #20834 (merged July 23, 2026), which introduced one `-lm` / `--load-mode` option with six values: `auto` (the default: memory-map unless the device does not support it), `none`, `mmap`, `mlock`, `mmap+mlock` and `dio`. Builds from September 9 onward reject the old flags, so launch scripts, systemd units and container commands that still pass them fail at startup.

▼ OPERATOR ANGLE

Search your launch scripts for the removed flags and translate them: `--no-mmap` becomes `--load-mode none`, `--mlock` becomes `--load-mode mlock`, mmap plus mlock becomes `--load-mode mmap+mlock`, and `--direct-io` becomes `--load-mode dio`. An older pinned build still accepts the old flags, but pin it deliberately. Wrappers that pass user-supplied extra arguments to `llama-server` are where this usually breaks.

SOURCE: https://github.com/ggml-org/llama.cpp/pull/28334[GITHUB-RELEASE]

▼ ENTITIES REFERENCED

TOOLllama.cpp

[pulse item] · runlocalai.co/pulse/llama-cpp-load-mode-replaces-mmap-mlock-flags