server: (router) lazy-load startup_models after main setup

tools/server/server-models.cpp

@@ -685,11 +685,10 @@ void server_models::load_models() {
                 models_to_load.size(), base_params.models_max));
         }
 
+        // to be lazy-loaded after main() setup phase is completed
+        startup_models = std::move(models_to_load);
+
         lk.unlock();
-        for (const auto & name : models_to_load) {
-            SRV_INF("(startup) loading model %s\n", name.c_str());
-            load(name);
-        }
     } else {
         // RELOAD: diff the new preset list against the current mapping and reconcile
         is_reloading = true;
@@ -877,6 +876,19 @@ void server_models::load_models() {
     }
 }
 
+void server_models::load_startup_models() {
+    std::vector<std::string> to_load;
+    {
+        std::lock_guard<std::mutex> lk(mutex);
+        to_load = std::move(startup_models);
+        startup_models.clear();
+    }
+    for (const auto & name : 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) {
     std::lock_guard<std::mutex> lk(mutex);
     auto it = mapping.find(name);

tools/server/server-models.h

@@ -136,6 +136,9 @@ private:
     // if true, the next get_meta() will trigger a reload of model list
     bool need_reload = false;
 
+    // models marked with load-on-startup
+    std::vector<std::string> startup_models;
+
     // conv_id -> model name that currently serves its stream session, lets the resumable stream
     // routes go straight to the owning child instead of polling every one. populated when
     // proxy_request forwards a POST carrying an X-Conversation-Id. best effort: a stale entry just
@@ -231,6 +234,9 @@ public:
     //   - if a model is not running, it will be added or updated according to the source
     void load_models();
 
+    // lazy-load startup_models, to be called after main() setup phase
+    void load_startup_models();
+
     // check if a model instance exists (thread-safe)
     bool has_model(const std::string & name);
 

tools/server/server.cpp

@@ -423,6 +423,18 @@ int llama_server(common_params & params, int argc, char ** argv) {
             ctx_http.stop();
         };
 
+        try {
+            models_routes->models.load_startup_models();
+        } catch (const std::exception & e) {
+            SRV_ERR("failed to load models on startup: %s\n", e.what());
+            ctx_http.stop();
+            if (ctx_http.thread.joinable()) {
+                ctx_http.thread.join();
+            }
+            clean_up();
+            return 1;
+        }
+
     } else {
         // setup clean up function, to be called before exit
         clean_up = [&ctx_http, &ctx_server, &mcp_mgr]() {