server: support load model on startup, support preset-only options (#18206)

* server: support autoload model, support preset-only options

* add docs

* load-on-startup

* fix

* Update common/arg.cpp

Co-authored-by: Pascal <admin@serveurperso.com>

---------

Co-authored-by: Pascal <admin@serveurperso.com>

tools/server/README.md

@@ -1480,6 +1480,9 @@ The precedence rule for preset options is as follows:
 2. **Model-specific options** defined in the preset file (e.g. `[ggml-org/MY-MODEL...]`)
 3. **Global options** defined in the preset file (`[*]`)
 
+We also offer additional options that are exclusive to presets (these aren't treated as command-line arguments):
+- `load-on-startup` (boolean): Controls whether the model loads automatically when the server starts
+
 ### Routing requests
 
 Requests are routed according to the requested model name.

tools/server/server-models.cpp

@@ -226,6 +226,26 @@ void server_models::load_models() {
             SRV_INF("  %c %s\n", has_custom ? '*' : ' ', name.c_str());
         }
     }
+
+    // load any autoload models
+    std::vector<std::string> models_to_load;
+    for (const auto & [name, inst] : mapping) {
+        std::string val;
+        if (inst.meta.preset.get_option(COMMON_ARG_PRESET_LOAD_ON_STARTUP, val)) {
+            models_to_load.push_back(name);
+        }
+    }
+    if ((int)models_to_load.size() > base_params.models_max) {
+        throw std::runtime_error(string_format(
+            "number of models to load on startup (%zu) exceeds models_max (%d)",
+            models_to_load.size(),
+            base_params.models_max
+        ));
+    }
+    for (const auto & name : models_to_load) {
+        SRV_INF("(startup) loading model %s\n", name.c_str());
+        load(name);
+    }
 }
 
 void server_models::update_meta(const std::string & name, const server_model_meta & meta) {

tools/server/server-models.h

@@ -103,27 +103,29 @@ public:
 
     void load_models();
 
-    // check if a model instance exists
+    // check if a model instance exists (thread-safe)
     bool has_model(const std::string & name);
 
-    // return a copy of model metadata
+    // return a copy of model metadata (thread-safe)
     std::optional<server_model_meta> get_meta(const std::string & name);
 
-    // return a copy of all model metadata
+    // return a copy of all model metadata (thread-safe)
     std::vector<server_model_meta> get_all_meta();
 
+    // load and unload model instances
+    // these functions are thread-safe
     void load(const std::string & name);
     void unload(const std::string & name);
     void unload_all();
 
-    // update the status of a model instance
+    // update the status of a model instance (thread-safe)
     void update_status(const std::string & name, server_model_status status);
 
-    // wait until the model instance is fully loaded
+    // wait until the model instance is fully loaded (thread-safe)
     // return when the model is loaded or failed to load
     void wait_until_loaded(const std::string & name);
 
-    // load the model if not loaded, otherwise do nothing
+    // load the model if not loaded, otherwise do nothing (thread-safe)
     // return false if model is already loaded; return true otherwise (meta may need to be refreshed)
     bool ensure_model_loaded(const std::string & name);