xinference 0.16.3__py3-none-any.whl → 1.0.1__py3-none-any.whl

xinference/model/embedding/core.py

@@ -21,6 +21,7 @@ from typing import Dict, List, Literal, Optional, Tuple, Union, no_type_check
 import numpy as np
 import torch
 
+from ..._compat import ROOT_KEY, ErrorWrapper, ValidationError
 from ...device_utils import empty_cache
 from ...types import Embedding, EmbeddingData, EmbeddingUsage
 from ..core import CacheableModelSpec, ModelDescription
@@ -193,6 +194,27 @@ class EmbeddingModel:
                 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(
@@ -202,11 +224,192 @@ class EmbeddingModel:
                 trust_remote_code=True,
             )
 
+    def _fix_langchain_openai_inputs(self, sentences: Union[str, List[str]]):
+        # Check if sentences is a two-dimensional list of integers
+        if (
+            isinstance(sentences, list)
+            and len(sentences) > 0
+            and isinstance(sentences[0], list)
+            and len(sentences[0]) > 0
+            and isinstance(sentences[0][0], int)
+        ):
+            # List[List[int]] stands for encoded inputs
+            import tiktoken
+
+            enc = tiktoken.get_encoding("cl100k_base")
+            lines_decoded = []
+
+            for line in sentences:
+                try:
+                    # Decode each token into bytes, then join them into a complete string
+                    output = b"".join(
+                        enc.decode_single_token_bytes(token) for token in line
+                    )
+                    # Convert the byte sequence into a UTF-8 encoded string
+                    decoded_line = output.decode("utf-8")
+                    lines_decoded.append(decoded_line)
+                except (ValueError, TypeError, UnicodeDecodeError) as e:
+                    raise ValidationError([ErrorWrapper(e, loc=ROOT_KEY)], self)
+
+            # Update sentences to be the list of decoded strings
+            if len(lines_decoded) == 1:
+                sentences = lines_decoded[0]
+            else:
+                sentences = lines_decoded
+        return sentences
+
     def create_embedding(self, sentences: Union[str, List[str]], **kwargs):
+        sentences = self._fix_langchain_openai_inputs(sentences)
+
+        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 +593,10 @@ class EmbeddingModel:
                 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 +608,30 @@ class EmbeddingModel:
             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 +653,38 @@ class EmbeddingModel:
 
         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,

xinference/model/embedding/model_spec.json

@@ -233,7 +233,7 @@
   },
   {
     "model_name": "gte-Qwen2",
-    "dimensions": 4096,
+    "dimensions": 3584,
     "max_tokens": 32000,
     "language": ["zh", "en"],
     "model_id": "Alibaba-NLP/gte-Qwen2-7B-instruct",

xinference/model/embedding/model_spec_modelscope.json

@@ -235,7 +235,7 @@
   },
   {
     "model_name": "gte-Qwen2",
-    "dimensions": 4096,
+    "dimensions": 3584,
     "max_tokens": 32000,
     "language": ["zh", "en"],
     "model_id": "iic/gte_Qwen2-7B-instruct",

xinference/model/image/stable_diffusion/core.py

@@ -17,9 +17,11 @@ import gc
 import inspect
 import itertools
 import logging
+import os
 import re
 import sys
 import warnings
+from glob import glob
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
 
 import PIL.Image
@@ -194,8 +196,9 @@ class DiffusionModel(SDAPIDiffusionModelMixin):
         if sys.platform != "darwin" and torch_dtype is None:
             # The following params crashes on Mac M2
             self._torch_dtype = self._kwargs["torch_dtype"] = torch.float16
-            self._kwargs["variant"] = "fp16"
-            self._kwargs["use_safetensors"] = True
+            self._kwargs["use_safetensors"] = any(
+                glob(os.path.join(self._model_path, "*/*.safetensors"))
+            )
         if isinstance(torch_dtype, str):
             self._kwargs["torch_dtype"] = getattr(torch, torch_dtype)
 

xinference/model/llm/__init__.py

@@ -143,6 +143,7 @@ def _install():
     )
     from .transformers.deepseek_vl import DeepSeekVLChatModel
     from .transformers.glm4v import Glm4VModel
+    from .transformers.glm_edge_v import GlmEdgeVModel
     from .transformers.intern_vl import InternVLChatModel
     from .transformers.internlm2 import Internlm2PytorchChatModel
     from .transformers.minicpmv25 import MiniCPMV25Model
@@ -193,6 +194,7 @@ def _install():
             DeepSeekV2PytorchModel,
             DeepSeekV2PytorchChatModel,
             OptPytorchModel,
+            GlmEdgeVModel,
         ]
     )
     if OmniLMMModel:  # type: ignore