Fast loads on python and liblemonade

bindings/python/bergamot.cpp

@@ -40,7 +40,8 @@ class ServicePyAdapter {
       : service_(make_service(config)) {}
 
   std::shared_ptr<_Model> modelFromConfig(const std::string &config) {
-    return service_.createCompatibleModel(config);
+    auto parsedConfig = marian::bergamot::parseOptionsFromString(config);
+    return service_.createCompatibleModel(parsedConfig);
   }
 
   std::shared_ptr<_Model> modelFromConfigPath(const std::string &configPath) {
@@ -208,12 +209,18 @@ PYBIND11_MODULE(_bergamot, m) {
 
   py::class_<Service::Config>(m, "ServiceConfig")
       .def(py::init<>([](size_t numWorkers, bool cacheEnabled, size_t cacheSize,
-                         size_t cacheMutexBuckets) {
-             return Service::Config{numWorkers, cacheEnabled, cacheSize,
-                                    cacheMutexBuckets};
+                         size_t cacheMutexBuckets, std::string logging) {
+             Service::Config config;
+             config.numWorkers = numWorkers;
+             config.cacheEnabled = cacheEnabled;
+             config.cacheSize = cacheSize;
+             config.cacheMutexBuckets = cacheMutexBuckets;
+             config.logger.level = logging;
+             return config;
            }),
            py::arg("numWorkers") = 1, py::arg("cacheEnabled") = false,
-           py::arg("cacheSize") = 20000, py::arg("cacheMutexBuckets") = 1)
+           py::arg("cacheSize") = 20000, py::arg("cacheMutexBuckets") = 1,
+           py::arg("logLevel") = "off")
       .def_readwrite("numWorkers", &Service::Config::numWorkers)
       .def_readwrite("cacheEnabled", &Service::Config::cacheEnabled)
       .def_readwrite("cacheSize", &Service::Config::cacheSize)

bindings/python/cmds.py

@@ -44,6 +44,14 @@ def embed_subparser(key: str, subparsers: argparse._SubParsersAction):
             help="Number of worker threads to use to translate",
             default=4,
         )
+
+        translate.add_argument(
+            "--log-level",
+            type=str,
+            default="off",
+            help="Set verbosity level of logging: trace, debug, info, warn, err(or), critical, off",
+        )
+
         # Tweak response-options for quick HTML in out via commandline
         options = translate.add_argument_group("response-options")
         options.add_argument("--html", type=bool, default=False)
@@ -54,7 +62,7 @@ def embed_subparser(key: str, subparsers: argparse._SubParsersAction):
     def execute(args: argparse.Namespace):
         # Build service
 
-        config = ServiceConfig(numWorkers=args.num_workers)
+        config = ServiceConfig(numWorkers=args.num_workers, logLevel=args.log_level)
         service = Service(config)
 
         models = [

lemonade/lib/translator.cpp

@@ -24,16 +24,10 @@ void Translator::translate(std::string input, const std::string &source,
       // FIXME
       modelConfig->set("workspace", 128);
 
-      marian::timer::Timer bundleTimer;
-      marian::bergamot::MemoryBundle memoryBundle =
-          marian::bergamot::getMemoryBundleFromConfig(modelConfig);
-      std::cout << fmt::format("Bundle loading from disk {} seconds.\n",
-                               bundleTimer.elapsed());
       marian::timer::Timer timer;
-      model =
-          service_.createCompatibleModel(modelConfig, std::move(memoryBundle));
+      model = service_.createCompatibleModel(modelConfig);
 
-      std::cout << fmt::format("Model building from bundle took {} seconds.\n",
+      std::cerr << fmt::format("Model building from bundle took {} seconds.\n",
                                timer.elapsed());
       manager_.cacheModel(m.code, model);
     }