モデルの読み込みを遅延させるように変更

run.py

@@ -861,6 +861,7 @@ def engine_manifest():
     parser.add_argument("--enable_cancellable_synthesis", action="store_true")
     parser.add_argument("--enable_guided_synthesis", action="store_true")
     parser.add_argument("--init_processes", type=int, default=2)
+    parser.add_argument("--load_all_models", action="store_true")
 
     # 引数へcpu_num_threadsの指定がなければ、環境変数をロールします。
     # 環境変数にもない場合は、Noneのままとします。
@@ -880,6 +881,7 @@ def engine_manifest():
         runtime_dirs=args.runtime_dir,
         cpu_num_threads=cpu_num_threads,
         enable_mock=args.enable_mock,
+        load_all_models=args.load_all_models,
     )
     assert len(synthesis_engines) != 0, "音声合成エンジンがありません。"
     latest_core_version = str(max([LooseVersion(ver) for ver in synthesis_engines]))

voicevox_engine/synthesis_engine/core_wrapper.py

@@ -339,7 +339,7 @@ def __init__(
         use_gpu: bool,
         core_dir: Path,
         cpu_num_threads: int = 0,
-        load_all_models: bool = True,
+        load_all_models: bool = False,
     ) -> None:
 
         self.core = load_core(core_dir, use_gpu)
@@ -354,12 +354,17 @@ def __init__(
         self.exist_suppoted_devices = False
         self.exist_finalize = False
         exist_cpu_num_threads = False
-        # TODO: version 0.12 から追加された load_model, is_model_loaded 関数に対応するため、
-        #       self.exist_load_model, self.exist_is_model_loaded 変数を定義する
+        self.exist_load_model = False
+        self.exist_is_model_loaded = False
 
         if is_version_0_12_core_or_later(core_dir):
             model_type = "onnxruntime"
-            # TODO: self.exist_load_model, self.exist_is_model_loaded 両方を True にする
+            self.exist_load_model = True
+            self.exist_is_model_loaded = True
+            self.core.load_model.argtypes = (c_int,)
+            self.core.load_model.restype = c_bool
+            self.core.is_model_loaded.argtypes = (c_int,)
+            self.core.is_model_loaded.restype = c_bool
         else:
             model_type = check_core_type(core_dir)
         assert model_type is not None
@@ -411,6 +416,11 @@ def __init__(
                     raise Exception(self.core.last_error_message().decode("utf-8"))
         finally:
             os.chdir(cwd)
+
+    def _lazy_init(self, speaker_id: int):
+        if self.exist_load_model and self.exist_load_model:
+            if not self.is_model_loaded(speaker_id):
+                self.load_model(speaker_id)
 
     def metas(self) -> str:
         return self.core.metas().decode("utf-8")
@@ -421,6 +431,7 @@ def yukarin_s_forward(
         phoneme_list: np.ndarray,
         speaker_id: np.ndarray,
     ) -> np.ndarray:
+        self._lazy_init(speaker_id[0])
         output = np.zeros((length,), dtype=np.float32)
         success = self.core.yukarin_s_forward(
             c_int(length),
@@ -443,6 +454,7 @@ def yukarin_sa_forward(
         end_accent_phrase_list: np.ndarray,
         speaker_id: np.ndarray,
     ) -> np.ndarray:
+        self._lazy_init(speaker_id[0])
         output = np.empty(
             (
                 len(speaker_id),
@@ -473,6 +485,7 @@ def decode_forward(
         phoneme: np.ndarray,
         speaker_id: np.ndarray,
     ) -> np.ndarray:
+        self._lazy_init(speaker_id[0])
         output = np.empty((length * 256,), dtype=np.float32)
         success = self.core.decode_forward(
             c_int(length),
@@ -496,3 +509,13 @@ def finalize(self) -> None:
             self.core.finalize()
             return
         raise NameError
+
+    def load_model(self, speaker_id: int) -> bool:
+        if self.exist_load_model:
+            return self.core.load_model(c_int(speaker_id))
+        raise NameError
+
+    def is_model_loaded(self, speaker_id: int) -> bool:
+        if self.exist_is_model_loaded:
+            return self.core.is_model_loaded(c_int(speaker_id))
+        raise NameError

voicevox_engine/synthesis_engine/make_synthesis_engines.py

@@ -16,6 +16,7 @@ def make_synthesis_engines(
     runtime_dirs: Optional[List[Path]] = None,
     cpu_num_threads: Optional[int] = None,
     enable_mock: bool = True,
+    load_all_models: bool = False,
 ) -> Dict[str, SynthesisEngineBase]:
     """
     音声ライブラリをロードして、音声合成エンジンを生成
@@ -36,6 +37,8 @@ def make_synthesis_engines(
         Noneのとき、ライブラリ側の挙動により論理コア数の半分か、物理コア数が指定される
     enable_mock: bool, optional, default=True
         コア読み込みに失敗したとき、代わりにmockを使用するかどうか
+    load_all_models: bool, optional, default=False
+        起動時に全てのモデルを読み込むかどうか
     """
     if cpu_num_threads == 0 or cpu_num_threads is None:
         print(
@@ -69,7 +72,7 @@ def make_synthesis_engines(
     synthesis_engines = {}
     for core_dir in voicelib_dirs:
         try:
-            core = CoreWrapper(use_gpu, core_dir, cpu_num_threads)
+            core = CoreWrapper(use_gpu, core_dir, cpu_num_threads, load_all_models)
             metas = json.loads(core.metas())
             core_version = metas[0]["version"]
             if core_version in synthesis_engines: