third

xinference/model/embedding/core.py

@@ -816,8 +816,10 @@ def create_embedding_model_instance(
         model_name,
         model_engine,
     )
+    devices = devices or ["cpu"]
     # model class should be one of flag, fastembed, sentence_transformer
-    device = devices[0] if devices is not None else None
+    # 这种写法会不会有问题？
+    device = devices[0] if isinstance(devices, list) else devices
     model = embedding_cls(model_uid, model_path, model_spec, device, **kwargs)
     model_description = EmbeddingModelDescription(
         subpool_addr, devices, model_spec, model_path=model_path

xinference/model/embedding/flag/core.py

@@ -30,16 +30,8 @@ def __init__(
     def load(self):
         try:
             from FlagEmbedding import BGEM3FlagModel
-
-            # from sentence_transformers import SentenceTransformer
-            # if FlagEmbedding.__version__ < "1.3.3":
-            #     raise ValueError(
-            #         "The FlagEmbedding version must be greater than 1.3.3. "
-            #         "Please upgrade your version via `pip install -U FlagEmbedding` or refer to "
-            #         "https://github.com/FlagOpen/FlagEmbedding"
-            #     )
         except ImportError:
-            error_message = "Failed to import module 'FlagEmbedding'"
+            error_message = "Failed to import module 'BGEM3FlagModel'"
             installation_guide = [
                 "Please make sure 'FlagEmbedding' is installed. ",
                 "You can install it by `pip install FlagEmbedding`\n",

xinference/model/embedding/sentence_transformer/core.py

@@ -317,6 +317,40 @@ def encode(
                 convert_to_numpy=False,
                 **kwargs,
             )
+        elif "clip" in self._model_spec.model_name.lower():
+            import base64
+            import re
+            from io import BytesIO
+
+            from PIL import Image
+
+            def base64_to_image(base64_str: str) -> Image.Image:
+                # base64_data = re.sub("^data:image/.+;base64,", "", base64_str)
+                base64_data = base64_str.split(",", 1)[1]
+                byte_data = base64.b64decode(base64_data)
+                image_data = BytesIO(byte_data)
+                img = Image.open(image_data)
+                return img
+
+            objs: list[dict[str, str]] = []
+            for item in sentences:
+                if isinstance(item, dict):
+                    if item.get("text") is not None:
+                        objs.append(item["text"])
+                    elif item.get("image") is not None:
+                        if re.match(r"^data:image/.+;base64,", item["image"]):
+                            image = base64_to_image(item["image"])
+                            objs.append(image)
+                        else:
+                            objs.append(item["image"])
+                    else:
+                        logger.error("Please check the input data.")
+            all_embeddings, all_token_nums = encode(
+                self._model,
+                objs,
+                convert_to_numpy=False,
+                **self._kwargs,
+            )
         else:
             all_embeddings, all_token_nums = encode(
                 self._model,

xinference/model/embedding/tests/test_embedding_models.py

@@ -252,7 +252,7 @@ def test_model_from_modelscope():
 
     model_path = cache(TEST_MODEL_SPEC_FROM_MODELSCOPE)
     model, _ = create_embedding_model_instance(
-        "mock", "cuda", "mock", "bge-small-zh-v1.5", "sentence_transformer", model_path
+        "mock", None, "mock", "bge-small-zh-v1.5", "sentence_transformer", model_path
     )
     # input is a string
     input_text = "乱条犹未变初黄，倚得东风势便狂。解把飞花蒙日月，不知天地有清霜。"