xinference 1.7.1__py3-none-any.whl → 1.8.0__py3-none-any.whl

xinference/model/audio/custom.py

@@ -11,29 +11,26 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 import logging
-import os
-from threading import Lock
 from typing import Any, List, Optional
 
 from ..._compat import (
     ROOT_KEY,
     ErrorWrapper,
+    Literal,
     Protocol,
     StrBytes,
     ValidationError,
     load_str_bytes,
 )
-from ...constants import XINFERENCE_CACHE_DIR, XINFERENCE_MODEL_DIR
-from .core import AudioModelFamilyV1
+from ..custom import ModelRegistry
+from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
-UD_AUDIO_LOCK = Lock()
-
 
-class CustomAudioModelFamilyV1(AudioModelFamilyV1):
+class CustomAudioModelFamilyV2(AudioModelFamilyV2):
+    version: Literal[2] = 2
     model_id: Optional[str]  # type: ignore
     model_revision: Optional[str]  # type: ignore
     model_uri: Optional[str]
@@ -47,7 +44,7 @@ class CustomAudioModelFamilyV1(AudioModelFamilyV1):
         encoding: str = "utf8",
         proto: Protocol = None,
         allow_pickle: bool = False,
-    ) -> AudioModelFamilyV1:
+    ) -> AudioModelFamilyV2:
         # See source code of BaseModel.parse_raw
         try:
             obj = load_str_bytes(
@@ -61,7 +58,7 @@ class CustomAudioModelFamilyV1(AudioModelFamilyV1):
         except (ValueError, TypeError, UnicodeDecodeError) as e:
             raise ValidationError([ErrorWrapper(e, loc=ROOT_KEY)], cls)
 
-        audio_spec: AudioModelFamilyV1 = cls.parse_obj(obj)
+        audio_spec: AudioModelFamilyV2 = cls.parse_obj(obj)
 
         # check model_family
         if audio_spec.model_family is None:
@@ -72,78 +69,36 @@ class CustomAudioModelFamilyV1(AudioModelFamilyV1):
         return audio_spec
 
 
-UD_AUDIOS: List[CustomAudioModelFamilyV1] = []
+UD_AUDIOS: List[CustomAudioModelFamilyV2] = []
+
 
+class AudioModelRegistry(ModelRegistry):
+    model_type = "audio"
 
-def get_user_defined_audios() -> List[CustomAudioModelFamilyV1]:
-    with UD_AUDIO_LOCK:
-        return UD_AUDIOS.copy()
+    def __init__(self):
+        from . import BUILTIN_AUDIO_MODELS
 
+        super().__init__()
+        self.models = UD_AUDIOS
+        self.builtin_models = list(BUILTIN_AUDIO_MODELS.keys())
 
-def register_audio(model_spec: CustomAudioModelFamilyV1, persist: bool):
-    from ...constants import XINFERENCE_MODEL_DIR
-    from ..utils import is_valid_model_name, is_valid_model_uri
-    from . import BUILTIN_AUDIO_MODELS, MODELSCOPE_AUDIO_MODELS
 
-    if not is_valid_model_name(model_spec.model_name):
-        raise ValueError(f"Invalid model name {model_spec.model_name}.")
+def get_user_defined_audios() -> List[CustomAudioModelFamilyV2]:
+    from ..custom import RegistryManager
 
-    model_uri = model_spec.model_uri
-    if model_uri and not is_valid_model_uri(model_uri):
-        raise ValueError(f"Invalid model URI {model_uri}.")
+    registry = RegistryManager.get_registry("audio")
+    return registry.get_custom_models()
 
-    with UD_AUDIO_LOCK:
-        for model_name in (
-            list(BUILTIN_AUDIO_MODELS.keys())
-            + list(MODELSCOPE_AUDIO_MODELS.keys())
-            + [spec.model_name for spec in UD_AUDIOS]
-        ):
-            if model_spec.model_name == model_name:
-                raise ValueError(
-                    f"Model name conflicts with existing model {model_spec.model_name}"
-                )
 
-        UD_AUDIOS.append(model_spec)
+def register_audio(model_spec: CustomAudioModelFamilyV2, persist: bool):
+    from ..custom import RegistryManager
 
-    if persist:
-        persist_path = os.path.join(
-            XINFERENCE_MODEL_DIR, "audio", f"{model_spec.model_name}.json"
-        )
-        os.makedirs(os.path.dirname(persist_path), exist_ok=True)
-        with open(persist_path, mode="w") as fd:
-            fd.write(model_spec.json())
+    registry = RegistryManager.get_registry("audio")
+    registry.register(model_spec, persist)
 
 
 def unregister_audio(model_name: str, raise_error: bool = True):
-    with UD_AUDIO_LOCK:
-        model_spec = None
-        for i, f in enumerate(UD_AUDIOS):
-            if f.model_name == model_name:
-                model_spec = f
-                break
-        if model_spec:
-            UD_AUDIOS.remove(model_spec)
-
-            persist_path = os.path.join(
-                XINFERENCE_MODEL_DIR, "audio", f"{model_spec.model_name}.json"
-            )
-            if os.path.exists(persist_path):
-                os.remove(persist_path)
-
-            cache_dir = os.path.join(XINFERENCE_CACHE_DIR, model_spec.model_name)
-            if os.path.exists(cache_dir):
-                logger.warning(
-                    f"Remove the cache of user-defined model {model_spec.model_name}. "
-                    f"Cache directory: {cache_dir}"
-                )
-                if os.path.isdir(cache_dir):
-                    os.rmdir(cache_dir)
-                else:
-                    logger.warning(
-                        f"Cache directory is not a soft link, please remove it manually."
-                    )
-        else:
-            if raise_error:
-                raise ValueError(f"Model {model_name} not found")
-            else:
-                logger.warning(f"Custom audio model {model_name} not found")
+    from ..custom import RegistryManager
+
+    registry = RegistryManager.get_registry("audio")
+    registry.unregister(model_name, raise_error)

xinference/model/audio/f5tts.py

@@ -19,7 +19,7 @@ from io import BytesIO
 from typing import TYPE_CHECKING, Optional, Union
 
 if TYPE_CHECKING:
-    from .core import AudioModelFamilyV1
+    from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
@@ -29,10 +29,11 @@ class F5TTSModel:
         self,
         model_uid: str,
         model_path: str,
-        model_spec: "AudioModelFamilyV1",
+        model_spec: "AudioModelFamilyV2",
         device: Optional[str] = None,
         **kwargs,
     ):
+        self.model_family = model_spec
         self._model_uid = model_uid
         self._model_path = model_path
         self._model_spec = model_spec

xinference/model/audio/f5tts_mlx.py

@@ -24,7 +24,7 @@ import numpy as np
 from tqdm import tqdm
 
 if TYPE_CHECKING:
-    from .core import AudioModelFamilyV1
+    from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
@@ -34,10 +34,11 @@ class F5TTSMLXModel:
         self,
         model_uid: str,
         model_path: str,
-        model_spec: "AudioModelFamilyV1",
+        model_spec: "AudioModelFamilyV2",
         device: Optional[str] = None,
         **kwargs,
     ):
+        self.model_family = model_spec
         self._model_uid = model_uid
         self._model_path = model_path
         self._model_spec = model_spec

xinference/model/audio/fish_speech.py

@@ -23,7 +23,7 @@ import torch
 from ...device_utils import get_available_device, is_device_available
 
 if TYPE_CHECKING:
-    from .core import AudioModelFamilyV1
+    from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
@@ -48,10 +48,11 @@ class FishSpeechModel:
         self,
         model_uid: str,
         model_path: str,
-        model_spec: "AudioModelFamilyV1",
+        model_spec: "AudioModelFamilyV2",
         device: Optional[str] = None,
         **kwargs,
     ):
+        self.model_family = model_spec
         self._model_uid = model_uid
         self._model_path = model_path
         self._model_spec = model_spec

xinference/model/audio/funasr.py

@@ -19,7 +19,7 @@ from typing import TYPE_CHECKING, List, Optional
 from ...device_utils import get_available_device, is_device_available
 
 if TYPE_CHECKING:
-    from .core import AudioModelFamilyV1
+    from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
@@ -29,10 +29,11 @@ class FunASRModel:
         self,
         model_uid: str,
         model_path: str,
-        model_spec: "AudioModelFamilyV1",
+        model_spec: "AudioModelFamilyV2",
         device: Optional[str] = None,
         **kwargs,
     ):
+        self.model_family = model_spec
         self._model_uid = model_uid
         self._model_path = model_path
         self._model_spec = model_spec
@@ -100,7 +101,11 @@ class FunASRModel:
             if not is_device_available(self._device):
                 raise ValueError(f"Device {self._device} is not available!")
 
-        kwargs = self._model_spec.default_model_config.copy()
+        kwargs = (
+            self._model_spec.default_model_config.copy()
+            if getattr(self._model_spec, "default_model_config", None)
+            else {}
+        )
         kwargs.update(self._kwargs)
         logger.debug("Loading FunASR model with kwargs: %s", kwargs)
         self._model = AutoModel(model=self._model_path, device=self._device, **kwargs)
@@ -131,12 +136,20 @@ class FunASRModel:
         with tempfile.NamedTemporaryFile(buffering=0) as f:
             f.write(audio)
 
-            kw = self._model_spec.default_transcription_config.copy()  # type: ignore
+            kw = (
+                self._model_spec.default_transcription_config.copy()  # type: ignore
+                if getattr(self._model_spec, "default_transcription_config", None)
+                else {}
+            )
             kw.update(kwargs)
             logger.debug("Calling FunASR model with kwargs: %s", kw)
             result = self._model.generate(  # type: ignore
                 input=f.name, cache={}, language=language, **kw
             )
+            if not result or not isinstance(result, list):
+                raise RuntimeError(f"FunASR returned empty or invalid result: {result}")
+            if "text" not in result[0]:
+                raise RuntimeError(f"Missing 'text' field in result[0]: {result[0]}")
             text = rich_transcription_postprocess(result[0]["text"])
 
             if response_format == "json":

xinference/model/audio/kokoro.py

@@ -20,7 +20,7 @@ import numpy as np
 from ...device_utils import get_available_device, is_device_available
 
 if TYPE_CHECKING:
-    from .core import AudioModelFamilyV1
+    from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
@@ -30,10 +30,11 @@ class KokoroModel:
         self,
         model_uid: str,
         model_path: str,
-        model_spec: "AudioModelFamilyV1",
+        model_spec: "AudioModelFamilyV2",
         device: Optional[str] = None,
         **kwargs,
     ):
+        self.model_family = model_spec
         self._model_uid = model_uid
         self._model_path = model_path
         self._model_spec = model_spec

xinference/model/audio/megatts.py

@@ -17,7 +17,7 @@ from io import BytesIO
 from typing import TYPE_CHECKING, Optional
 
 if TYPE_CHECKING:
-    from .core import AudioModelFamilyV1
+    from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
@@ -27,10 +27,11 @@ class MegaTTSModel:
         self,
         model_uid: str,
         model_path: str,
-        model_spec: "AudioModelFamilyV1",
+        model_spec: "AudioModelFamilyV2",
         device: Optional[str] = None,
         **kwargs,
     ):
+        self.model_family = model_spec
         self._model_uid = model_uid
         self._model_path = model_path
         self._model_spec = model_spec

xinference/model/audio/melotts.py

@@ -18,7 +18,7 @@ from typing import TYPE_CHECKING, Optional
 from ...device_utils import get_available_device, is_device_available
 
 if TYPE_CHECKING:
-    from .core import AudioModelFamilyV1
+    from .core import AudioModelFamilyV2
 
 logger = logging.getLogger(__name__)
 
@@ -28,10 +28,11 @@ class MeloTTSModel:
         self,
         model_uid: str,
         model_path: str,
-        model_spec: "AudioModelFamilyV1",
+        model_spec: "AudioModelFamilyV2",
         device: Optional[str] = None,
         **kwargs,
     ):
+        self.model_family = model_spec
         self._model_uid = model_uid
         self._model_path = model_path
         self._model_spec = model_spec