Tools
llama-server can finally be configured entirely from a systemd EnvironmentFile
PR #27380 (merged 2026-09-21) adds `LLAMA_ARG_*` environment variables for `--temp`, `--top-p`, `--min-p`, `--repeat-penalty`, `--presence-penalty` and `--frequency-penalty`, closing the gap that forced sampling parameters into the command line. The contributor's motivation is concrete and worth stealing: running llama-server on Debian Forky as a systemd unit with `EnvironmentFile=/etc/default/llama-server` and a bare `ExecStart=/usr/bin/llama-server`, so the service definition never changes when a model's recommended sampling settings do. `--top-k` and other options are still command-line only.
Source
↳ Follow the thread