lemonade
This commit is contained in:
146
modules/nixos/services/lemonade/default.nix
Normal file
146
modules/nixos/services/lemonade/default.nix
Normal file
@@ -0,0 +1,146 @@
|
||||
{
|
||||
config,
|
||||
lib,
|
||||
pkgs,
|
||||
namespace,
|
||||
...
|
||||
}:
|
||||
let
|
||||
inherit (lib) mkForce getExe;
|
||||
inherit (lib.${namespace}) mkModule mkOpt;
|
||||
|
||||
name = "lemonade";
|
||||
cfg = config.${namespace}.services.${name};
|
||||
|
||||
# lemonade-server serve args built from config options
|
||||
serveArgs = lib.concatStringsSep " " (
|
||||
[
|
||||
"serve"
|
||||
"--no-tray"
|
||||
"--port ${toString cfg.port}"
|
||||
"--host ${cfg.host}"
|
||||
"--log-level ${cfg.logLevel}"
|
||||
]
|
||||
++ lib.optional (cfg.maxLoadedModels != 1) "--max-loaded-models ${toString cfg.maxLoadedModels}"
|
||||
++ lib.optional (cfg.extraModelsDir != null) "--extra-models-dir ${cfg.extraModelsDir}"
|
||||
++ cfg.extraArgs
|
||||
);
|
||||
|
||||
lemonadeConfig = mkModule {
|
||||
inherit config name;
|
||||
description = "Lemonade local LLM server";
|
||||
|
||||
options = {
|
||||
# Override mkModule's default port of 80 with lemonade's actual default.
|
||||
port = mkOpt lib.types.int 8000 "Port lemonade-router listens on";
|
||||
|
||||
host = mkOpt lib.types.str "127.0.0.1" "Address lemonade-router binds to";
|
||||
|
||||
logLevel = mkOpt (lib.types.enum [
|
||||
"critical"
|
||||
"error"
|
||||
"warning"
|
||||
"info"
|
||||
"debug"
|
||||
"trace"
|
||||
]) "info" "Log level for lemonade-router";
|
||||
|
||||
maxLoadedModels = mkOpt lib.types.int 1 "Maximum number of models to keep loaded simultaneously";
|
||||
|
||||
extraModelsDir =
|
||||
mkOpt (lib.types.nullOr lib.types.str) null
|
||||
"Extra directory scanned for local GGUF model files";
|
||||
|
||||
extraArgs =
|
||||
mkOpt (lib.types.listOf lib.types.str) [ ]
|
||||
"Extra arguments passed verbatim to lemonade-server serve";
|
||||
|
||||
modelsDir =
|
||||
mkOpt lib.types.str "/var/lib/${name}/models"
|
||||
"Directory where downloaded models are stored (exposed as HF_HOME)";
|
||||
|
||||
apiKeyFile =
|
||||
mkOpt (lib.types.nullOr lib.types.str) null
|
||||
"Path to a file containing the LEMONADE_API_KEY (e.g. a sops secret path)";
|
||||
};
|
||||
|
||||
moduleConfig = {
|
||||
# Install the package system-wide so lemonade-server / lemonade-router are
|
||||
# available in PATH for interactive use alongside the daemon.
|
||||
environment.systemPackages = [ pkgs.${namespace}.lemonade ];
|
||||
|
||||
systemd.services.${name} = {
|
||||
description = "Lemonade local LLM server";
|
||||
wantedBy = [ "multi-user.target" ];
|
||||
after = [
|
||||
"network.target"
|
||||
"network-online.target"
|
||||
];
|
||||
wants = [ "network-online.target" ];
|
||||
|
||||
# lemonade-server discover lemonade-router by reading /proc/self/exe,
|
||||
# so we must use ExecStart with the real binary, not a shell wrapper.
|
||||
serviceConfig = {
|
||||
Type = "simple";
|
||||
ExecStart = "${getExe pkgs.${namespace}.lemonade} ${serveArgs}";
|
||||
|
||||
User = name;
|
||||
Group = name;
|
||||
DynamicUser = mkForce false;
|
||||
|
||||
# Models and HuggingFace cache land under modelsDir.
|
||||
# HF_HOME overrides the default ~/.cache/huggingface location.
|
||||
Environment = [
|
||||
"HF_HOME=${cfg.modelsDir}"
|
||||
"XDG_RUNTIME_DIR=/run/${name}"
|
||||
];
|
||||
|
||||
# Load an API key from a secrets file if provided.
|
||||
EnvironmentFile = lib.optional (cfg.apiKeyFile != null) cfg.apiKeyFile;
|
||||
|
||||
# Runtime directory for PID file / lock file (created automatically
|
||||
# by systemd and owned by the service user).
|
||||
RuntimeDirectory = name;
|
||||
RuntimeDirectoryMode = "0755";
|
||||
|
||||
# Persistent state: models cache.
|
||||
StateDirectory = name;
|
||||
StateDirectoryMode = "0750";
|
||||
|
||||
# Home directory for the service user (needed by some HF tooling).
|
||||
WorkingDirectory = "/var/lib/${name}";
|
||||
|
||||
Restart = "on-failure";
|
||||
RestartSec = "5s";
|
||||
|
||||
StandardOutput = "journal";
|
||||
StandardError = "journal";
|
||||
SyslogIdentifier = name;
|
||||
|
||||
# Hardening — lemonade needs network access and subprocess execution
|
||||
# for spawning llama.cpp / whisper.cpp backends.
|
||||
NoNewPrivileges = true;
|
||||
PrivateTmp = true;
|
||||
ProtectSystem = "strict";
|
||||
ProtectHome = true;
|
||||
ReadWritePaths = [
|
||||
"/var/lib/${name}"
|
||||
"/run/${name}"
|
||||
];
|
||||
};
|
||||
};
|
||||
|
||||
users.users.${name} = {
|
||||
isSystemUser = true;
|
||||
group = name;
|
||||
home = "/var/lib/${name}";
|
||||
createHome = true;
|
||||
description = "Lemonade LLM server daemon";
|
||||
};
|
||||
users.groups.${name} = { };
|
||||
};
|
||||
};
|
||||
in
|
||||
{
|
||||
imports = [ lemonadeConfig ];
|
||||
}
|
||||
Reference in New Issue
Block a user