xinference 1.7.1.post1__py3-none-any.whl → 1.8.0__py3-none-any.whl

xinference/model/embedding/sentence_transformers/core.py

@@ -14,20 +14,16 @@
 
 import importlib.util
 import logging
-from collections import defaultdict
 from typing import List, Optional, Union, no_type_check
 
 import numpy as np
 import torch
 
-from ....types import Dict, Embedding, EmbeddingData, EmbeddingUsage
-from ..core import EmbeddingModel, EmbeddingModelSpec
+from ....device_utils import is_device_available
+from ....types import Embedding, EmbeddingData, EmbeddingUsage
+from ..core import EmbeddingModel, EmbeddingModelFamilyV2, EmbeddingSpecV1
 
 logger = logging.getLogger(__name__)
-
-# Used for check whether the model is cached.
-# Init when registering all the builtin models.
-MODEL_NAME_TO_REVISION: Dict[str, List[str]] = defaultdict(list)
 SENTENCE_TRANSFORMER_MODEL_LIST: List[str] = []
 
 
@@ -76,8 +72,8 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
                 torch_dtype = torch.float32
 
         if (
-            "gte" in self._model_spec.model_name.lower()
-            and "qwen2" in self._model_spec.model_name.lower()
+            "gte" in self.model_family.model_name.lower()
+            and "qwen2" in self.model_family.model_name.lower()
         ):
             model_kwargs = {"device_map": "auto"}
             if torch_dtype:
@@ -87,10 +83,12 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
                 device=self._device,
                 model_kwargs=model_kwargs,
             )
-        elif "qwen3" in self._model_spec.model_name.lower():
+        elif "qwen3" in self.model_family.model_name.lower():
             # qwen3 embedding
             flash_attn_installed = importlib.util.find_spec("flash_attn") is not None
-            flash_attn_enabled = self._kwargs.get("enable_flash_attn", True)
+            flash_attn_enabled = self._kwargs.get(
+                "enable_flash_attn", is_device_available("cuda")
+            )
             model_kwargs = {"device_map": "auto"}
             tokenizer_kwargs = {}
             if flash_attn_installed and flash_attn_enabled:
@@ -119,7 +117,8 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
                 trust_remote_code=True,
             )
 
-        self._tokenizer = self._model.tokenizer
+        if hasattr(self._model, "tokenizer"):
+            self._tokenizer = self._model.tokenizer
 
     def create_embedding(
         self,
@@ -227,8 +226,8 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
                 device = model._target_device
 
             if (
-                "gte" in self._model_spec.model_name.lower()
-                and "qwen2" in self._model_spec.model_name.lower()
+                "gte" in self.model_family.model_name.lower()
+                and "qwen2" in self.model_family.model_name.lower()
             ):
                 model.to(device)
 
@@ -254,7 +253,10 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
                 features.update(extra_features)
                 # when batching, the attention mask 1 means there is a token
                 # thus we just sum up it to get the total number of tokens
-                if "clip" in self._model_spec.model_name.lower():
+                if (
+                    "clip" in self.model_family.model_name.lower()
+                    or "jina-embeddings-v4" in self.model_family.model_name.lower()
+                ):
                     if "input_ids" in features and hasattr(
                         features["input_ids"], "numel"
                     ):
@@ -322,8 +324,8 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
 
         # seems already support prompt in embedding model
         if (
-            "gte" in self._model_spec.model_name.lower()
-            and "qwen2" in self._model_spec.model_name.lower()
+            "gte" in self.model_family.model_name.lower()
+            and "qwen2" in self.model_family.model_name.lower()
         ):
             all_embeddings, all_token_nums = encode(
                 self._model,
@@ -332,7 +334,10 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
                 convert_to_numpy=False,
                 **kwargs,
             )
-        elif "clip" in self._model_spec.model_name.lower():
+        elif (
+            "clip" in self.model_family.model_name.lower()
+            or "jina-embeddings-v4" in self.model_family.model_name.lower()
+        ):
             import base64
             import re
             from io import BytesIO
@@ -409,6 +414,11 @@ class SentenceTransformerEmbeddingModel(EmbeddingModel):
         return importlib.util.find_spec("sentence_transformers") is not None
 
     @classmethod
-    def match_json(cls, model_spec: EmbeddingModelSpec) -> bool:
+    def match_json(
+        cls,
+        model_family: EmbeddingModelFamilyV2,
+        model_spec: EmbeddingSpecV1,
+        quantization: str,
+    ) -> bool:
         # As default embedding engine, sentence-transformer support all models
-        return True
+        return model_spec.model_format in ["pytorch"]

xinference/model/embedding/vllm/core.py

@@ -17,7 +17,7 @@ import logging
 from typing import List, Union
 
 from ....types import Embedding, EmbeddingData, EmbeddingUsage
-from ..core import EmbeddingModel, EmbeddingModelSpec
+from ..core import EmbeddingModel, EmbeddingModelFamilyV2, EmbeddingSpecV1
 
 logger = logging.getLogger(__name__)
 SUPPORTED_MODELS_PREFIXES = ["bge", "gte", "text2vec", "m3e", "gte", "Qwen3"]
@@ -88,8 +88,14 @@ class VLLMEmbeddingModel(EmbeddingModel):
         return importlib.util.find_spec("vllm") is not None
 
     @classmethod
-    def match_json(cls, model_spec: EmbeddingModelSpec) -> bool:
-        prefix = model_spec.model_name.split("-", 1)[0]
-        if prefix in SUPPORTED_MODELS_PREFIXES:
-            return True
+    def match_json(
+        cls,
+        model_family: EmbeddingModelFamilyV2,
+        model_spec: EmbeddingSpecV1,
+        quantization: str,
+    ) -> bool:
+        if model_spec.model_format in ["pytorch"]:
+            prefix = model_family.model_name.split("-", 1)[0]
+            if prefix in SUPPORTED_MODELS_PREFIXES:
+                return True
         return False

xinference/model/flexible/__init__.py

@@ -11,7 +11,6 @@
 # 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 codecs
 import json
 import logging
@@ -25,6 +24,8 @@ from .core import (
     FlexibleModelSpec,
     generate_flexible_model_description,
     get_flexible_model_descriptions,
+)
+from .custom import (
     get_flexible_models,
     register_flexible_model,
     unregister_flexible_model,
@@ -34,7 +35,12 @@ logger = logging.getLogger(__name__)
 
 
 def register_custom_model():
-    model_dir = os.path.join(XINFERENCE_MODEL_DIR, "flexible")
+    from ..custom import migrate_from_v1_to_v2
+
+    # migrate from v1 to v2 first
+    migrate_from_v1_to_v2("flexible", FlexibleModelSpec)
+
+    model_dir = os.path.join(XINFERENCE_MODEL_DIR, "v2", "flexible")
     if os.path.isdir(model_dir):
         for f in os.listdir(model_dir):
             try:

xinference/model/flexible/core.py

@@ -14,21 +14,19 @@
 
 import json
 import logging
-import os
 from collections import defaultdict
-from threading import Lock
-from typing import Dict, List, Optional, Tuple
+from typing import Dict, List, Optional
 
-from ...constants import XINFERENCE_CACHE_DIR, XINFERENCE_MODEL_DIR
-from ..core import CacheableModelSpec, ModelDescription, VirtualEnvSettings
+from ..._compat import Literal
+from ..core import CacheableModelSpec, VirtualEnvSettings
+from ..utils import ModelInstanceInfoMixin
 from .utils import get_launcher
 
 logger = logging.getLogger(__name__)
 
-FLEXIBLE_MODEL_LOCK = Lock()
 
-
-class FlexibleModelSpec(CacheableModelSpec):
+class FlexibleModelSpec(CacheableModelSpec, ModelInstanceInfoMixin):
+    version: Literal[1, 2] = 2
     model_id: Optional[str]  # type: ignore
     model_description: Optional[str]
     model_uri: Optional[str]
@@ -39,42 +37,26 @@ class FlexibleModelSpec(CacheableModelSpec):
     def parser_args(self):
         return json.loads(self.launcher_args)
 
+    class Config:
+        extra = "allow"
 
-class FlexibleModelDescription(ModelDescription):
-    def __init__(
-        self,
-        address: Optional[str],
-        devices: Optional[List[str]],
-        model_spec: FlexibleModelSpec,
-        model_path: Optional[str] = None,
-    ):
-        super().__init__(address, devices, model_path=model_path)
-        self._model_spec = model_spec
-
-    @property
-    def spec(self):
-        return self._model_spec
-
-    def to_dict(self):
+    def to_description(self):
         return {
             "model_type": "flexible",
-            "address": self.address,
-            "accelerators": self.devices,
-            "model_name": self._model_spec.model_name,
-            "launcher": self._model_spec.launcher,
-            "launcher_args": self._model_spec.launcher_args,
+            "address": getattr(self, "address", None),
+            "accelerators": getattr(self, "accelerators", None),
+            "model_name": self.model_name,
+            "launcher": self.launcher,
+            "launcher_args": self.launcher_args,
         }
 
-    def get_model_version(self) -> str:
-        return f"{self._model_spec.model_name}"
-
     def to_version_info(self):
         return {
-            "model_version": self.get_model_version(),
+            "model_version": self.model_name,
             "cache_status": True,
-            "model_file_location": self._model_spec.model_uri,
-            "launcher": self._model_spec.launcher,
-            "launcher_args": self._model_spec.launcher_args,
+            "model_file_location": self.model_uri,
+            "launcher": self.launcher,
+            "launcher_args": self.launcher_args,
         }
 
 
@@ -82,9 +64,7 @@ def generate_flexible_model_description(
     model_spec: FlexibleModelSpec,
 ) -> Dict[str, List[Dict]]:
     res = defaultdict(list)
-    res[model_spec.model_name].append(
-        FlexibleModelDescription(None, None, model_spec).to_version_info()
-    )
+    res[model_spec.model_name].append(model_spec.to_version_info())
     return res
 
 
@@ -92,93 +72,22 @@ FLEXIBLE_MODELS: List[FlexibleModelSpec] = []
 FLEXIBLE_MODEL_DESCRIPTIONS: Dict[str, List[Dict]] = defaultdict(list)
 
 
-def get_flexible_models():
-    with FLEXIBLE_MODEL_LOCK:
-        return FLEXIBLE_MODELS.copy()
-
-
 def get_flexible_model_descriptions():
     import copy
 
     return copy.deepcopy(FLEXIBLE_MODEL_DESCRIPTIONS)
 
 
-def register_flexible_model(model_spec: FlexibleModelSpec, persist: bool):
-    from ..utils import is_valid_model_name, is_valid_model_uri
-
-    if not is_valid_model_name(model_spec.model_name):
-        raise ValueError(f"Invalid model name {model_spec.model_name}.")
-
-    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}.")
-
-    if model_spec.launcher_args:
-        try:
-            model_spec.parser_args()
-        except Exception:
-            raise ValueError(f"Invalid model launcher args {model_spec.launcher_args}.")
-
-    with FLEXIBLE_MODEL_LOCK:
-        for model_name in [spec.model_name for spec in FLEXIBLE_MODELS]:
-            if model_spec.model_name == model_name:
-                raise ValueError(
-                    f"Model name conflicts with existing model {model_spec.model_name}"
-                )
-        FLEXIBLE_MODELS.append(model_spec)
-
-    if persist:
-        persist_path = os.path.join(
-            XINFERENCE_MODEL_DIR, "flexible", 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())
-
-
-def unregister_flexible_model(model_name: str, raise_error: bool = True):
-    with FLEXIBLE_MODEL_LOCK:
-        model_spec = None
-        for i, f in enumerate(FLEXIBLE_MODELS):
-            if f.model_name == model_name:
-                model_spec = f
-                break
-        if model_spec:
-            FLEXIBLE_MODELS.remove(model_spec)
-
-            persist_path = os.path.join(
-                XINFERENCE_MODEL_DIR, "flexible", 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.islink(cache_dir):
-                    os.remove(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"Model {model_name} not found")
-
-
 class FlexibleModel:
     def __init__(
         self,
         model_uid: str,
         model_path: str,
+        model_family: FlexibleModelSpec,
         device: Optional[str] = None,
         config: Optional[Dict] = None,
     ):
+        self.model_family = model_family
         self._model_uid = model_uid
         self._model_path = model_path
         self._device = device
@@ -213,19 +122,20 @@ class FlexibleModel:
 
 
 def match_flexible_model(model_name):
+    from .custom import get_flexible_models
+
     for model_spec in get_flexible_models():
         if model_name == model_spec.model_name:
             return model_spec
+    return None
 
 
 def create_flexible_model_instance(
-    subpool_addr: str,
-    devices: List[str],
     model_uid: str,
     model_name: str,
     model_path: Optional[str] = None,
     **kwargs,
-) -> Tuple[FlexibleModel, FlexibleModelDescription]:
+) -> FlexibleModel:
     model_spec = match_flexible_model(model_name)
     if not model_path:
         model_path = model_spec.model_uri
@@ -237,7 +147,4 @@ def create_flexible_model_instance(
         model_uid=model_uid, model_spec=model_spec, **kwargs
     )
 
-    model_description = FlexibleModelDescription(
-        subpool_addr, devices, model_spec, model_path=model_path
-    )
-    return model, model_description
+    return model

xinference/model/flexible/custom.py

@@ -0,0 +1,69 @@
+from typing import TYPE_CHECKING
+
+from ..custom import ModelRegistry
+
+if TYPE_CHECKING:
+    from .core import FlexibleModelSpec
+
+
+class FlexibleModelRegistry(ModelRegistry):
+    model_type = "flexible"
+
+    def __init__(self):
+        from .core import FLEXIBLE_MODELS
+
+        super().__init__()
+        self.models = FLEXIBLE_MODELS
+        self.builtin_models = []
+
+    def register(self, model_spec: "FlexibleModelSpec", persist: bool):
+        from ..cache_manager import CacheManager
+        from ..utils import is_valid_model_name, is_valid_model_uri
+
+        if not is_valid_model_name(model_spec.model_name):
+            raise ValueError(f"Invalid model name {model_spec.model_name}.")
+
+        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}.")
+
+        if model_spec.launcher_args:
+            try:
+                model_spec.parser_args()
+            except Exception:
+                raise ValueError(
+                    f"Invalid model launcher args {model_spec.launcher_args}."
+                )
+
+        with self.lock:
+            for model_name in [spec.model_name for spec in self.models]:
+                if model_spec.model_name == model_name:
+                    raise ValueError(
+                        f"Model name conflicts with existing model {model_spec.model_name}"
+                    )
+            self.models.append(model_spec)
+
+        if persist:
+            cache_manager = CacheManager(model_spec)
+            cache_manager.register_custom_model(self.model_type)
+
+
+def get_flexible_models():
+    from ..custom import RegistryManager
+
+    registry = RegistryManager.get_registry("flexible")
+    return registry.get_custom_models()
+
+
+def register_flexible_model(model_spec: "FlexibleModelSpec", persist: bool):
+    from ..custom import RegistryManager
+
+    registry = RegistryManager.get_registry("flexible")
+    registry.register(model_spec, persist)
+
+
+def unregister_flexible_model(model_name: str, raise_error: bool = True):
+    from ..custom import RegistryManager
+
+    registry = RegistryManager.get_registry("flexible")
+    registry.unregister(model_name, raise_error)

xinference/model/flexible/launchers/image_process_launcher.py

@@ -63,6 +63,7 @@ def launcher(model_uid: str, model_spec: FlexibleModelSpec, **kwargs) -> Flexibl
         return ImageRemoveBackgroundModel(
             model_uid=model_uid,
             model_path=model_spec.model_uri,  # type: ignore
+            model_family=model_spec,
             device=device,
             config=kwargs,
         )

xinference/model/flexible/launchers/modelscope_launcher.py

@@ -43,5 +43,9 @@ def launcher(model_uid: str, model_spec: FlexibleModelSpec, **kwargs) -> Flexibl
         raise ValueError("model_path required")
 
     return ModelScopePipelineModel(
-        model_uid=model_uid, model_path=model_path, device=device, config=kwargs
+        model_uid=model_uid,
+        model_path=model_path,
+        model_family=model_spec,
+        device=device,
+        config=kwargs,
     )

xinference/model/flexible/launchers/transformers_launcher.py

@@ -51,13 +51,25 @@ def launcher(model_uid: str, model_spec: FlexibleModelSpec, **kwargs) -> Flexibl
 
     if task == "text-classification":
         return TransformersTextClassificationModel(
-            model_uid=model_uid, model_path=model_path, device=device, config=kwargs
+            model_uid=model_uid,
+            model_path=model_path,
+            model_family=model_spec,
+            device=device,
+            config=kwargs,
         )
     elif task == "mock":
         return MockModel(
-            model_uid=model_uid, model_path=model_path, device=device, config=kwargs
+            model_uid=model_uid,
+            model_path=model_path,
+            model_family=model_spec,
+            device=device,
+            config=kwargs,
         )
     else:
         return AutoModel(
-            model_uid=model_uid, model_path=model_path, device=device, config=kwargs
+            model_uid=model_uid,
+            model_path=model_path,
+            model_family=model_spec,
+            device=device,
+            config=kwargs,
         )

xinference/model/flexible/launchers/yolo_launcher.py

@@ -58,5 +58,9 @@ def launcher(model_uid: str, model_spec: FlexibleModelSpec, **kwargs) -> Flexibl
         raise ValueError("model_path required")
 
     return UltralyticsModel(
-        model_uid=model_uid, model_path=model_path, device=device, config=kwargs
+        model_uid=model_uid,
+        model_path=model_path,
+        model_family=model_spec,
+        device=device,
+        config=kwargs,
     )

xinference/model/image/__init__.py

@@ -16,20 +16,17 @@ import codecs
 import json
 import os
 import warnings
-from itertools import chain
 
+from ..utils import flatten_model_src
 from .core import (
     BUILTIN_IMAGE_MODELS,
     IMAGE_MODEL_DESCRIPTIONS,
-    MODEL_NAME_TO_REVISION,
-    MODELSCOPE_IMAGE_MODELS,
-    ImageModelFamilyV1,
+    ImageModelFamilyV2,
     generate_image_description,
-    get_cache_status,
     get_image_model_descriptions,
 )
 from .custom import (
-    CustomImageModelFamilyV1,
+    CustomImageModelFamilyV2,
     get_user_defined_images,
     register_image,
     unregister_image,
@@ -38,15 +35,19 @@ from .custom import (
 
 def register_custom_model():
     from ...constants import XINFERENCE_MODEL_DIR
+    from ..custom import migrate_from_v1_to_v2
 
-    user_defined_image_dir = os.path.join(XINFERENCE_MODEL_DIR, "image")
+    # migrate from v1 to v2 first
+    migrate_from_v1_to_v2("image", CustomImageModelFamilyV2)
+
+    user_defined_image_dir = os.path.join(XINFERENCE_MODEL_DIR, "v2", "image")
     if os.path.isdir(user_defined_image_dir):
         for f in os.listdir(user_defined_image_dir):
             try:
                 with codecs.open(
                     os.path.join(user_defined_image_dir, f), encoding="utf-8"
                 ) as fd:
-                    user_defined_image_family = CustomImageModelFamilyV1.parse_obj(
+                    user_defined_image_family = CustomImageModelFamilyV2.parse_obj(
                         json.load(fd)
                     )
                     register_image(user_defined_image_family, persist=False)
@@ -56,12 +57,10 @@ def register_custom_model():
 
 def _install():
     load_model_family_from_json("model_spec.json", BUILTIN_IMAGE_MODELS)
-    load_model_family_from_json("model_spec_modelscope.json", MODELSCOPE_IMAGE_MODELS)
 
     # register model description
-    for model_name, model_spec in chain(
-        MODELSCOPE_IMAGE_MODELS.items(), BUILTIN_IMAGE_MODELS.items()
-    ):
+    for model_name, model_specs in BUILTIN_IMAGE_MODELS.items():
+        model_spec = [x for x in model_specs if x.model_hub == "huggingface"][0]
         IMAGE_MODEL_DESCRIPTIONS.update(generate_image_description(model_spec))
 
     register_custom_model()
@@ -72,13 +71,14 @@ def _install():
 
 def load_model_family_from_json(json_filename, target_families):
     json_path = os.path.join(os.path.dirname(__file__), json_filename)
-    target_families.update(
-        dict(
-            (spec["model_name"], ImageModelFamilyV1(**spec))
-            for spec in json.load(codecs.open(json_path, "r", encoding="utf-8"))
-        )
-    )
-    for model_name, model_spec in target_families.items():
-        MODEL_NAME_TO_REVISION[model_name].append(model_spec.model_revision)
+    flattened_model_specs = []
+    for spec in json.load(codecs.open(json_path, "r", encoding="utf-8")):
+        flattened_model_specs.extend(flatten_model_src(spec))
+
+    for spec in flattened_model_specs:
+        if spec["model_name"] not in target_families:
+            target_families[spec["model_name"]] = [ImageModelFamilyV2(**spec)]
+        else:
+            target_families[spec["model_name"]].append(ImageModelFamilyV2(**spec))
 
     del json_path

xinference/model/image/cache_manager.py

@@ -0,0 +1,62 @@
+import os
+from typing import Optional
+
+from ..cache_manager import CacheManager
+
+
+class ImageCacheManager(CacheManager):
+    def cache_gguf(self, quantization: Optional[str] = None):
+        from ..utils import IS_NEW_HUGGINGFACE_HUB, retry_download, symlink_local_file
+        from .core import ImageModelFamilyV2
+
+        if not quantization:
+            return None
+
+        assert isinstance(self._model_family, ImageModelFamilyV2)
+        cache_dir = self.get_cache_dir()
+
+        if not self._model_family.gguf_model_file_name_template:
+            raise NotImplementedError(
+                f"{self._model_family.model_name} does not support GGUF quantization"
+            )
+        if quantization not in (self._model_family.gguf_quantizations or []):
+            raise ValueError(
+                f"Cannot support quantization {quantization}, "
+                f"available quantizations: {self._model_family.gguf_quantizations}"
+            )
+
+        filename = self._model_family.gguf_model_file_name_template.format(quantization=quantization)  # type: ignore
+        full_path = os.path.join(cache_dir, filename)
+
+        if self._model_family.model_hub == "huggingface":
+            import huggingface_hub
+
+            use_symlinks = {}
+            if not IS_NEW_HUGGINGFACE_HUB:
+                use_symlinks = {"local_dir_use_symlinks": True, "local_dir": cache_dir}
+            download_file_path = retry_download(
+                huggingface_hub.hf_hub_download,
+                self._model_family.model_name,
+                None,
+                self._model_family.gguf_model_id,
+                filename=filename,
+                **use_symlinks,
+            )
+            if IS_NEW_HUGGINGFACE_HUB:
+                symlink_local_file(download_file_path, cache_dir, filename)
+        elif self._model_family.model_hub == "modelscope":
+            from modelscope.hub.file_download import model_file_download
+
+            download_file_path = retry_download(
+                model_file_download,
+                self._model_family.model_name,
+                None,
+                self._model_family.gguf_model_id,
+                filename,
+                revision=self._model_family.model_revision,
+            )
+            symlink_local_file(download_file_path, cache_dir, filename)
+        else:
+            raise NotImplementedError
+
+        return full_path