FEAT: create_embedding add field model_replica

xinference/api/restful_api.py

@@ -1331,6 +1331,7 @@ async def create_embedding(self, request: Request) -> Response:
 
         try:
             embedding = await model.create_embedding(body.input, **kwargs)
+            embedding["model"] = model_uid  # type: ignore
             return Response(embedding, media_type="application/json")
         except Exception as e:
             e = await self._get_model_last_error(model.uid, e)

xinference/model/embedding/core.py

@@ -693,7 +693,8 @@ def base64_to_image(base64_str: str) -> Image.Image:
                 if not is_bge_m3_flag_model and not kwargs.get("return_sparse")
                 else "dict"
             ),
-            model=self._model_uid,
+            model="",
+            model_replica=self._model_uid,
             data=embedding_list,
             usage=usage,
         )

xinference/types.py

@@ -78,6 +78,7 @@ class EmbeddingData(TypedDict):
 class Embedding(TypedDict):
     object: Literal["list"]
     model: str
+    model_replica: str
     data: List[EmbeddingData]
     usage: EmbeddingUsage