FEAT: support sparse vector for bge-m3 (#2540)

Co-authored-by: pengjunfeng11 <179464367@qq.com>
Co-authored-by: Junfeng Peng <pengjunfeng@hyperchain.cn>

xinference/api/restful_api.py

@@ -489,6 +489,16 @@ async def internal_exception_handler(request: Request, exc: Exception):
                 else None
             ),
         )
+        self._router.add_api_route(
+            "/v1/convert_ids_to_tokens",
+            self.convert_ids_to_tokens,
+            methods=["POST"],
+            dependencies=(
+                [Security(self._auth_service, scopes=["models:read"])]
+                if self.is_authenticated()
+                else None
+            ),
+        )
         self._router.add_api_route(
             "/v1/rerank",
             self.rerank,
@@ -1312,6 +1322,41 @@ async def create_embedding(self, request: Request) -> Response:
             await self._report_error_event(model_uid, str(e))
             raise HTTPException(status_code=500, detail=str(e))
 
+    async def convert_ids_to_tokens(self, request: Request) -> Response:
+        payload = await request.json()
+        body = CreateEmbeddingRequest.parse_obj(payload)
+        model_uid = body.model
+        exclude = {
+            "model",
+            "input",
+            "user",
+        }
+        kwargs = {key: value for key, value in payload.items() if key not in exclude}
+
+        try:
+            model = await (await self._get_supervisor_ref()).get_model(model_uid)
+        except ValueError as ve:
+            logger.error(str(ve), exc_info=True)
+            await self._report_error_event(model_uid, str(ve))
+            raise HTTPException(status_code=400, detail=str(ve))
+        except Exception as e:
+            logger.error(e, exc_info=True)
+            await self._report_error_event(model_uid, str(e))
+            raise HTTPException(status_code=500, detail=str(e))
+
+        try:
+            decoded_texts = await model.convert_ids_to_tokens(body.input, **kwargs)
+            return Response(decoded_texts, media_type="application/json")
+        except RuntimeError as re:
+            logger.error(re, exc_info=True)
+            await self._report_error_event(model_uid, str(re))
+            self.handle_request_limit_error(re)
+            raise HTTPException(status_code=400, detail=str(re))
+        except Exception as e:
+            logger.error(e, exc_info=True)
+            await self._report_error_event(model_uid, str(e))
+            raise HTTPException(status_code=500, detail=str(e))
+
     async def rerank(self, request: Request) -> Response:
         payload = await request.json()
         body = RerankRequest.parse_obj(payload)

xinference/client/restful/restful_client.py

@@ -126,6 +126,43 @@ def create_embedding(self, input: Union[str, List[str]], **kwargs) -> "Embedding
         response_data = response.json()
         return response_data
 
+    def convert_ids_to_tokens(
+        self, input: Union[List, List[List]], **kwargs
+    ) -> List[str]:
+        """
+        Convert token IDs to human readable tokens via RESTful APIs.
+
+        Parameters
+        ----------
+        input: Union[List, List[List]]
+            Input token IDs to convert, can be a single list of token IDs or a list of token ID lists.
+            To convert multiple sequences in a single request, pass a list of token ID lists.
+
+        Returns
+        -------
+        list
+            A list of decoded tokens in human readable format.
+
+        Raises
+        ------
+        RuntimeError
+            Report the failure of token conversion and provide the error message.
+
+        """
+        url = f"{self._base_url}/v1/convert_ids_to_tokens"
+        request_body = {
+            "model": self._model_uid,
+            "input": input,
+        }
+        request_body.update(kwargs)
+        response = requests.post(url, json=request_body, headers=self.auth_headers)
+        if response.status_code != 200:
+            raise RuntimeError(
+                f"Failed to decode token ids, detail: {_get_error_string(response)}"
+            )
+        response_data = response.json()
+        return response_data
+
 
 class RESTfulRerankModelHandle(RESTfulModelHandle):
     def rerank(

xinference/core/model.py

@@ -794,6 +794,19 @@ async def create_embedding(self, input: Union[str, List[str]], *args, **kwargs):
             f"Model {self._model.model_spec} is not for creating embedding."
         )
 
+    @request_limit
+    @log_async(logger=logger)
+    async def convert_ids_to_tokens(
+        self, input: Union[List, List[List]], *args, **kwargs
+    ):
+        kwargs.pop("request_id", None)
+        if hasattr(self._model, "convert_ids_to_tokens"):
+            return await self._call_wrapper_json(
+                self._model.convert_ids_to_tokens, input, *args, **kwargs
+            )
+
+        raise AttributeError(f"Model {self._model.model_spec} can convert token id.")
+
     @request_limit
     @log_async(logger=logger)
     async def rerank(

xinference/model/embedding/core.py

@@ -193,6 +193,27 @@ def to(self, *args, **kwargs):
                 device=self._device,
                 model_kwargs=model_kwargs,
             )
+        elif (
+            self._kwargs.get("hybrid_mode")
+            and "m3" in self._model_spec.model_name.lower()
+        ):
+            try:
+                from FlagEmbedding import BGEM3FlagModel
+            except ImportError:
+                error_message = "Failed to import module 'BGEM3FlagModel'"
+                installation_guide = [
+                    "Please make sure 'FlagEmbedding' is installed. ",
+                    "You can install it by `pip install FlagEmbedding`\n",
+                ]
+                raise ImportError(f"{error_message}\n\n{''.join(installation_guide)}")
+
+            model_kwargs = {"torch_dtype": torch_dtype} if torch_dtype else None
+            self._model = BGEM3FlagModel(
+                self._model_path,
+                device=self._device,
+                model_kwargs=model_kwargs,
+                trust_remote_code=True,
+            )
         else:
             model_kwargs = {"torch_dtype": torch_dtype} if torch_dtype else None
             self._model = SentenceTransformer(
@@ -203,10 +224,155 @@ def to(self, *args, **kwargs):
             )
 
     def create_embedding(self, sentences: Union[str, List[str]], **kwargs):
+        from FlagEmbedding import BGEM3FlagModel
         from sentence_transformers import SentenceTransformer
 
         kwargs.setdefault("normalize_embeddings", True)
 
+        @no_type_check
+        def _encode_bgem3(
+            model: Union[SentenceTransformer, BGEM3FlagModel],
+            sentences: Union[str, List[str]],
+            batch_size: int = 32,
+            show_progress_bar: bool = None,
+            output_value: str = "sparse_embedding",
+            convert_to_numpy: bool = True,
+            convert_to_tensor: bool = False,
+            device: str = None,
+            normalize_embeddings: bool = False,
+            **kwargs,
+        ):
+            """
+            Computes sentence embeddings with bge-m3 model
+            Nothing special here, just replace sentence-transformer with FlagEmbedding
+            TODO: think about how to solve the redundant code of encode method in the future
+
+            :param sentences: the sentences to embed
+            :param batch_size: the batch size used for the computation
+            :param show_progress_bar: Output a progress bar when encode sentences
+            :param output_value:  Default sentence_embedding, to get sentence embeddings. Can be set to token_embeddings to get wordpiece token embeddings. Set to None, to get all output values
+            :param convert_to_numpy: If true, the output is a list of numpy vectors. Else, it is a list of pytorch tensors.
+            :param convert_to_tensor: If true, you get one large tensor as return. Overwrites any setting from convert_to_numpy
+            :param device: Which torch.device to use for the computation
+            :param normalize_embeddings: If set to true, returned vectors will have length 1. In that case, the faster dot-product (util.dot_score) instead of cosine similarity can be used.
+
+            :return:
+               By default, a list of tensors is returned. If convert_to_tensor, a stacked tensor is returned. If convert_to_numpy, a numpy matrix is returned.
+            """
+            import torch
+            from tqdm.autonotebook import trange
+
+            if show_progress_bar is None:
+                show_progress_bar = (
+                    logger.getEffectiveLevel() == logging.INFO
+                    or logger.getEffectiveLevel() == logging.DEBUG
+                )
+
+            if convert_to_tensor:
+                convert_to_numpy = False
+
+            if output_value != "sparse_embedding":
+                convert_to_tensor = False
+                convert_to_numpy = False
+
+            input_was_string = False
+            if isinstance(sentences, str) or not hasattr(
+                sentences, "__len__"
+            ):  # Cast an individual sentence to a list with length 1
+                sentences = [sentences]
+                input_was_string = True
+
+            if device is None:
+                # Same as SentenceTransformer.py
+                from sentence_transformers.util import get_device_name
+
+                device = get_device_name()
+                logger.info(f"Use pytorch device_name: {device}")
+
+            all_embeddings = []
+            all_token_nums = 0
+
+            # The original code does not support other inference engines
+            def _text_length(text):
+                if isinstance(text, dict):  # {key: value} case
+                    return len(next(iter(text.values())))
+                elif not hasattr(text, "__len__"):  # Object has no len() method
+                    return 1
+                elif len(text) == 0 or isinstance(
+                    text[0], int
+                ):  # Empty string or list of ints
+                    return len(text)
+                else:
+                    return sum(
+                        [len(t) for t in text]
+                    )  # Sum of length of individual strings
+
+            length_sorted_idx = np.argsort([-_text_length(sen) for sen in sentences])
+            sentences_sorted = [sentences[idx] for idx in length_sorted_idx]
+
+            for start_index in trange(
+                0,
+                len(sentences),
+                batch_size,
+                desc="Batches",
+                disable=not show_progress_bar,
+            ):
+                sentences_batch = sentences_sorted[
+                    start_index : start_index + batch_size
+                ]
+
+                with torch.no_grad():
+                    out_features = model.encode(sentences_batch, **kwargs)
+
+                    if output_value == "token_embeddings":
+                        embeddings = []
+                        for token_emb, attention in zip(
+                            out_features[output_value], out_features["attention_mask"]
+                        ):
+                            last_mask_id = len(attention) - 1
+                            while (
+                                last_mask_id > 0 and attention[last_mask_id].item() == 0
+                            ):
+                                last_mask_id -= 1
+
+                            embeddings.append(token_emb[0 : last_mask_id + 1])
+                    elif output_value is None:  # Return all outputs
+                        embeddings = []
+                        for sent_idx in range(len(out_features["sentence_embedding"])):
+                            row = {
+                                name: out_features[name][sent_idx]
+                                for name in out_features
+                            }
+                            embeddings.append(row)
+                    # for sparse embedding
+                    else:
+                        if kwargs.get("return_sparse"):
+                            embeddings = out_features["lexical_weights"]
+                        else:
+                            embeddings = out_features["dense_vecs"]
+
+                        if convert_to_numpy:
+                            embeddings = embeddings.cpu()
+
+                    all_embeddings.extend(embeddings)
+
+            all_embeddings = [
+                all_embeddings[idx] for idx in np.argsort(length_sorted_idx)
+            ]
+
+            if convert_to_tensor:
+                if len(all_embeddings):
+                    all_embeddings = torch.stack(all_embeddings)
+                else:
+                    all_embeddings = torch.Tensor()
+            elif convert_to_numpy:
+                all_embeddings = np.asarray([emb.numpy() for emb in all_embeddings])
+
+            if input_was_string:
+                all_embeddings = all_embeddings[0]
+
+            return all_embeddings, all_token_nums
+
         # copied from sentence-transformers, and modify it to return tokens num
         @no_type_check
         def encode(
@@ -390,6 +556,10 @@ def encode(
                 convert_to_numpy=False,
                 **kwargs,
             )
+        elif isinstance(self._model, BGEM3FlagModel):
+            all_embeddings, all_token_nums = _encode_bgem3(
+                self._model, sentences, convert_to_numpy=False, **kwargs
+            )
         else:
             all_embeddings, all_token_nums = encode(
                 self._model,
@@ -401,14 +571,30 @@ def encode(
             all_embeddings = [all_embeddings]
         embedding_list = []
         for index, data in enumerate(all_embeddings):
-            embedding_list.append(
-                EmbeddingData(index=index, object="embedding", embedding=data.tolist())
-            )
+            if kwargs.get("return_sparse") and isinstance(self._model, BGEM3FlagModel):
+                embedding_list.append(
+                    EmbeddingData(
+                        index=index,
+                        object="embedding",
+                        embedding={k: float(v) for k, v in data.items()},
+                    )
+                )
+            else:
+                embedding_list.append(
+                    EmbeddingData(
+                        index=index, object="embedding", embedding=data.tolist()
+                    )
+                )
         usage = EmbeddingUsage(
             prompt_tokens=all_token_nums, total_tokens=all_token_nums
         )
         result = Embedding(
-            object="list",
+            object=(
+                "list"  # type: ignore
+                if not isinstance(self._model, BGEM3FlagModel)
+                and not kwargs.get("return_sparse")
+                else "dict"
+            ),
             model=self._model_uid,
             data=embedding_list,
             usage=usage,
@@ -430,6 +616,38 @@ def encode(
 
         return result
 
+    def convert_ids_to_tokens(
+        self,
+        batch_token_ids: Union[List[Union[int, str]], List[List[Union[int, str]]]],
+        **kwargs,
+    ) -> Union[List[str]]:
+        batch_decoded_texts: List[str] = []
+
+        assert self._model is not None
+
+        if isinstance(batch_token_ids, (int, str)):
+            return self._model.tokenizer.convert_ids_to_tokens(
+                [int(str(batch_token_ids))]
+            )[0]
+
+        # check if it's a nested list
+        if (
+            isinstance(batch_token_ids, list)
+            and batch_token_ids
+            and isinstance(batch_token_ids[0], list)
+        ):
+            for token_ids in batch_token_ids:
+                token_ids = [int(token_id) for token_id in token_ids]
+                batch_decoded_texts.append(
+                    self._model.tokenizer.convert_ids_to_tokens(token_ids)
+                )
+        else:
+            batch_token_ids = [int(token_id) for token_id in batch_token_ids]
+            batch_decoded_texts = self._model.tokenizer.convert_ids_to_tokens(
+                batch_token_ids
+            )
+        return batch_decoded_texts
+
 
 def match_embedding(
     model_name: str,