xinference 0.15.2__py3-none-any.whl → 0.15.4__py3-none-any.whl

xinference/model/audio/model_spec.json

@@ -71,6 +71,14 @@
     "model_ability": "audio-to-text",
     "multilingual": true
   },
+  {
+    "model_name": "whisper-large-v3-turbo",
+    "model_family": "whisper",
+    "model_id": "openai/whisper-large-v3-turbo",
+    "model_revision": "41f01f3fe87f28c78e2fbf8b568835947dd65ed9",
+    "model_ability": "audio-to-text",
+    "multilingual": true
+  },
   {
     "model_name": "Belle-distilwhisper-large-v2-zh",
     "model_family": "whisper",

xinference/model/audio/model_spec_modelscope.json

@@ -8,6 +8,15 @@
     "model_ability": "audio-to-text",
     "multilingual": true
   },
+  {
+    "model_name": "whisper-large-v3-turbo",
+    "model_family": "whisper",
+    "model_hub": "modelscope",
+    "model_id": "AI-ModelScope/whisper-large-v3-turbo",
+    "model_revision": "master",
+    "model_ability": "audio-to-text",
+    "multilingual": true
+  },
   {
     "model_name": "SenseVoiceSmall",
     "model_family": "funasr",

xinference/model/embedding/core.py

@@ -141,7 +141,15 @@ class EmbeddingModel:
 
     def load(self):
         try:
+            import sentence_transformers
             from sentence_transformers import SentenceTransformer
+
+            if sentence_transformers.__version__ < "3.1.0":
+                raise ValueError(
+                    "The sentence_transformers version must be greater than 3.1.0. "
+                    "Please upgrade your version via `pip install -U sentence_transformers` or refer to "
+                    "https://github.com/UKPLab/sentence-transformers"
+                )
         except ImportError:
             error_message = "Failed to import module 'SentenceTransformer'"
             installation_guide = [
@@ -173,9 +181,6 @@ class EmbeddingModel:
                 )
                 torch_dtype = torch.float32
 
-        from ..utils import patch_trust_remote_code
-
-        patch_trust_remote_code()
         if (
             "gte" in self._model_spec.model_name.lower()
             and "qwen2" in self._model_spec.model_name.lower()
@@ -191,7 +196,10 @@ class EmbeddingModel:
         else:
             model_kwargs = {"torch_dtype": torch_dtype} if torch_dtype else None
             self._model = SentenceTransformer(
-                self._model_path, device=self._device, model_kwargs=model_kwargs
+                self._model_path,
+                device=self._device,
+                model_kwargs=model_kwargs,
+                trust_remote_code=True,
             )
 
     def create_embedding(self, sentences: Union[str, List[str]], **kwargs):
@@ -213,6 +221,7 @@ class EmbeddingModel:
             convert_to_tensor: bool = False,
             device: str = None,
             normalize_embeddings: bool = False,
+            **kwargs,
         ):
             """
             Computes sentence embeddings
@@ -317,7 +326,7 @@ class EmbeddingModel:
                 all_token_nums += features["attention_mask"].sum().item()
 
                 with torch.no_grad():
-                    out_features = model.forward(features)
+                    out_features = model.forward(features, **kwargs)
 
                     if output_value == "token_embeddings":
                         embeddings = []

xinference/model/embedding/model_spec.json

@@ -238,5 +238,12 @@
     "language": ["zh", "en"],
     "model_id": "Alibaba-NLP/gte-Qwen2-7B-instruct",
     "model_revision": "e26182b2122f4435e8b3ebecbf363990f409b45b"
+  },
+  {
+    "model_name": "jina-embeddings-v3",
+    "dimensions": 1024,
+    "max_tokens": 8192,
+    "language": ["zh", "en"],
+    "model_id": "jinaai/jina-embeddings-v3"
   }
 ]

xinference/model/embedding/model_spec_modelscope.json

@@ -233,12 +233,20 @@
     "model_id": "AI-ModelScope/m3e-large",
     "model_hub": "modelscope"
   },
-    {
+  {
     "model_name": "gte-Qwen2",
     "dimensions": 4096,
     "max_tokens": 32000,
     "language": ["zh", "en"],
     "model_id": "iic/gte_Qwen2-7B-instruct",
     "model_hub": "modelscope"
+  },
+  {
+    "model_name": "jina-embeddings-v3",
+    "dimensions": 1024,
+    "max_tokens": 8192,
+    "language": ["zh", "en"],
+    "model_id": "jinaai/jina-embeddings-v3",
+    "model_hub": "modelscope"
   }
 ]

xinference/model/image/core.py

@@ -23,8 +23,6 @@ from ..core import CacheableModelSpec, ModelDescription
 from ..utils import valid_model_revision
 from .stable_diffusion.core import DiffusionModel
 
-MAX_ATTEMPTS = 3
-
 logger = logging.getLogger(__name__)
 
 MODEL_NAME_TO_REVISION: Dict[str, List[str]] = defaultdict(list)
@@ -210,18 +208,19 @@ def create_image_model_instance(
         for name in controlnet:
             for cn_model_spec in model_spec.controlnet:
                 if cn_model_spec.model_name == name:
-                    if not model_path:
-                        model_path = cache(cn_model_spec)
-                    controlnet_model_paths.append(model_path)
+                    controlnet_model_path = cache(cn_model_spec)
+                    controlnet_model_paths.append(controlnet_model_path)
                     break
             else:
                 raise ValueError(
                     f"controlnet `{name}` is not supported for model `{model_name}`."
                 )
         if len(controlnet_model_paths) == 1:
-            kwargs["controlnet"] = controlnet_model_paths[0]
+            kwargs["controlnet"] = (controlnet[0], controlnet_model_paths[0])
         else:
-            kwargs["controlnet"] = controlnet_model_paths
+            kwargs["controlnet"] = [
+                (n, path) for n, path in zip(controlnet, controlnet_model_paths)
+            ]
     if not model_path:
         model_path = cache(model_spec)
     if peft_model_config is not None:

xinference/model/image/sdapi.py

@@ -11,11 +11,12 @@
 # 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 base64
 import io
 import warnings
 
-from PIL import Image
+from PIL import Image, ImageOps
 
 
 class SDAPIToDiffusersConverter:
@@ -26,11 +27,12 @@ class SDAPIToDiffusersConverter:
         "width",
         "height",
         "sampler_name",
+        "progressor",
     }
     txt2img_arg_mapping = {
         "steps": "num_inference_steps",
         "cfg_scale": "guidance_scale",
-        # "denoising_strength": "strength",
+        "denoising_strength": "strength",
     }
     img2img_identical_args = {
         "prompt",
@@ -39,12 +41,15 @@ class SDAPIToDiffusersConverter:
         "width",
         "height",
         "sampler_name",
+        "progressor",
     }
     img2img_arg_mapping = {
         "init_images": "image",
+        "mask": "mask_image",
         "steps": "num_inference_steps",
         "cfg_scale": "guidance_scale",
         "denoising_strength": "strength",
+        "inpaint_full_res_padding": "padding_mask_crop",
     }
 
     @staticmethod
@@ -121,12 +126,38 @@ class SDAPIDiffusionModelMixin:
 
     def img2img(self, **kwargs):
         init_images = kwargs.pop("init_images", [])
-        kwargs["init_images"] = [self._decode_b64_img(i) for i in init_images]
+        kwargs["init_images"] = init_images = [
+            self._decode_b64_img(i) for i in init_images
+        ]
+        if len(init_images) == 1:
+            kwargs["init_images"] = init_images[0]
+        mask_image = kwargs.pop("mask", None)
+        if mask_image:
+            if kwargs.pop("inpainting_mask_invert"):
+                mask_image = ImageOps.invert(mask_image)
+
+            kwargs["mask"] = self._decode_b64_img(mask_image)
+
+            # process inpaint_full_res and inpaint_full_res_padding
+            if kwargs.pop("inpaint_full_res", None):
+                kwargs["inpaint_full_res_padding"] = kwargs.pop(
+                    "inpaint_full_res_padding", 0
+                )
+            else:
+                # inpaint_full_res_padding is turned `into padding_mask_crop`
+                # in diffusers, if padding_mask_crop is passed, it will do inpaint_full_res
+                # so if not inpaint_full_rs, we need to pop this option
+                kwargs.pop("inpaint_full_res_padding", None)
+
         clip_skip = kwargs.get("override_settings", {}).get("clip_skip")
         converted_kwargs = self._check_kwargs("img2img", kwargs)
         if clip_skip:
             converted_kwargs["clip_skip"] = clip_skip
-        result = self.image_to_image(response_format="b64_json", **converted_kwargs)  # type: ignore
+
+        if not converted_kwargs.get("mask_image"):
+            result = self.image_to_image(response_format="b64_json", **converted_kwargs)  # type: ignore
+        else:
+            result = self.inpainting(response_format="b64_json", **converted_kwargs)  # type: ignore
 
         # convert to SD API result
         return {