slop: make sure per-model settings are properly loaded

Author: meh

tools/server/server-model-manager.cpp

@@ -232,8 +232,12 @@ void server_model_manager::unload(const std::string& name, server_context& ctx)
     info.status = SERVER_MODEL_STATUS_UNLOADED;
     info.last_used = 0;
 
-    // Call server_context's unload
-    ctx.unload_current_model();
+    // Only destroy the model in ctx_server if this model is the one currently loaded.
+    // Without this guard, unloading any model marked LOADED would destroy whatever
+    // ctx_server happens to hold (which may be a different model).
+    if (ctx.has_model_loaded() && ctx.get_current_model_path() == info.model_path) {
+        ctx.unload_current_model();
+    }
 
     cv_.notify_all();
 }

tools/server/server.cpp

@@ -405,7 +405,11 @@ int main(int argc, char ** argv) {
             base_info.tags = params.model_tags;
             base_info.status = SERVER_MODEL_STATUS_LOADED;
             base_info.last_used = ggml_time_ms();
-            base_info.preset = cli_load_result.base_preset;
+            // Args after the first -- are this model's per-model preset; args before -- are
+            // global defaults already in `params`, so on reload we re-apply only this preset.
+            base_info.preset = cli_load_result.model_presets.empty()
+                ? cli_load_result.base_preset
+                : cli_load_result.model_presets[0];
             if (!base_info.name.empty()) {
                 model_manager->add_model(std::move(base_info));
             }
@@ -506,6 +510,17 @@ int main(int argc, char ** argv) {
             params.model.path = cli_base_model_path;
         }
 
+        // Build the first model's load params: global defaults (`params`, args before --)
+        // overlaid with its own per-model preset (args in the first -- block). Each model is
+        // independent — keep `params` unmutated so other models don't inherit this one's args.
+        common_params first_model_params = params;
+        if (cli_has_model_presets && !cli_load_result.model_presets.empty()) {
+            cli_load_result.model_presets[0].apply_to_params(first_model_params);
+            if (!cli_base_model_path.empty()) {
+                first_model_params.model.path = cli_base_model_path;
+            }
+        }
+
         // load the model
         LOG_INF("%s: loading model\n", __func__);
 
@@ -515,7 +530,7 @@ int main(int argc, char ** argv) {
             });
         }
 
-        if (!ctx_server.load_model(params)) {
+        if (!ctx_server.load_model(first_model_params)) {
             clean_up();
             if (ctx_http.thread.joinable()) {
                 ctx_http.thread.join();
@@ -872,7 +887,7 @@ int main(int argc, char ** argv) {
                 return res;
             }));
 
-            ctx_http.post("/models/unload", ex_wrapper([model_mgr2 = model_manager.get(), &ctx_server, model_manager_base_model_name](const server_http_req & req) -> server_http_res_ptr {
+            ctx_http.post("/models/unload", ex_wrapper([model_mgr2 = model_manager.get(), &ctx_server](const server_http_req & req) -> server_http_res_ptr {
                 auto res = std::make_unique<server_http_res>();
                 json body = json::parse(req.body);
                 std::string name = json_value(body, "model", std::string());
@@ -889,13 +904,6 @@ int main(int argc, char ** argv) {
                     return res;
                 }
 
-                // Don't allow unloading the base model
-                if (name == model_manager_base_model_name) {
-                    res->status = 400;
-                    res->data = safe_json_to_str({{"error", format_error_response("base model cannot be unloaded", ERROR_TYPE_INVALID_REQUEST)}});
-                    return res;
-                }
-
                 model_mgr2->unload(name, ctx_server);
                 res_ok(res, {{"success", true}});
                 return res;