server: fix remote preset handling, add test (#24938)

* server: add test for remote preset

* fix remote preset handling

* fix

* fix test

tools/server/server-models.cpp

@@ -224,7 +224,7 @@ void server_model_meta::update_caps() {
         });
         params.offline = true;
         // params.skip_download = true; // TODO: ideally, we should validate the model here, but it takes too much time
-        common_params_handle_models(params, LLAMA_EXAMPLE_SERVER);
+        common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, {});
         if (params.mmproj.path.empty()) {
             multimodal = { false, false };
         } else {
@@ -1393,7 +1393,9 @@ struct server_download_state : public common_download_callback {
 
     bool run(common_params & params) {
         try {
-            common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, this);
+            common_params_handle_models_params p;
+            p.callback = this;
+            common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, p);
             is_ok = true;
         } catch (const std::exception & e) {
             auto model_name = params.model.get_name();

tools/server/server.cpp

@@ -89,6 +89,17 @@ int llama_server(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
+    // note: router mode also accepts -hf remote-preset, so we need to check that first
+    if (!params.model.hf_repo.empty()) {
+        try {
+            common_params_handle_models_params handle_params;
+            handle_params.preset_only = true;
+            common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, handle_params);
+        } catch (const std::exception & e) {
+            // ignored for now
+        }
+    }
+
     // router server never loads a model and must not touch the GPU
     const bool is_router_server = params.model.path.empty()
                                && params.model.hf_repo.empty();
@@ -263,7 +274,7 @@ int llama_server(int argc, char ** argv) {
         return child.run_download(params);
     } else if (!is_router_server) {
         // single-model mode (NOT spawned by router)
-        common_params_handle_models(params, LLAMA_EXAMPLE_SERVER);
+        common_params_handle_models(params, LLAMA_EXAMPLE_SERVER, {});
     }
 
     //

tools/server/tests/unit/test_router.py

@@ -256,6 +256,25 @@ def test_router_reload_models():
         os.remove(preset_path)
 
 
+def test_router_remote_preset():
+    global server
+    server.model_hf_repo = "ggml-org/test-preset-ci"
+    server.model_hf_file = None
+    server.offline = False
+    server.start()
+
+    # Should see preset models in GET /models
+    res = server.make_request("GET", "/models")
+    assert res.status_code == 200
+    ids = {item["id"] for item in res.body.get("data", [])}
+    assert "tinygemma3-preset" in ids
+    assert "stories260K-test" in ids
+
+    # Should be able to load a preset model
+    model_id = "tinygemma3-preset"
+    _load_model_and_wait(model_id)
+
+
 MODEL_DOWNLOAD_ID = "ggml-org/test-model-router-download:F16"
 MODEL_DOWNLOAD_TIMEOUT = 30