xinference 0.9.4__py3-none-any.whl → 0.10.0__py3-none-any.whl

xinference/model/llm/pytorch/deepseek_vl.py

@@ -0,0 +1,232 @@
+# Copyright 2022-2023 XProbe Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 logging
+import os.path
+import tempfile
+import time
+import uuid
+from concurrent.futures import ThreadPoolExecutor
+from io import BytesIO
+from typing import Any, Dict, Iterator, List, Optional, Tuple, Union
+
+import requests
+import torch
+
+from ....model.utils import select_device
+from ....types import (
+    ChatCompletion,
+    ChatCompletionChoice,
+    ChatCompletionChunk,
+    ChatCompletionMessage,
+    CompletionUsage,
+)
+from ..llm_family import LLMFamilyV1, LLMSpecV1
+from .core import PytorchChatModel, PytorchGenerateConfig
+
+logger = logging.getLogger(__name__)
+
+
+class DeepSeekVLChatModel(PytorchChatModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._tokenizer = None
+        self._model = None
+        self._vl_chat_processor = None
+        self._type = None
+
+    @classmethod
+    def match(
+        cls, model_family: "LLMFamilyV1", model_spec: "LLMSpecV1", quantization: str
+    ) -> bool:
+        if "deepseek" in model_family.model_name:
+            return True
+        return False
+
+    def load(self):
+        from transformers import AutoModelForCausalLM
+
+        from ....thirdparty.deepseek_vl.models import (
+            MultiModalityCausalLM,
+            VLChatProcessor,
+        )
+
+        self._device = self._pytorch_model_config.get("device", "auto")
+        self._device = select_device(self._device)
+        self._type = torch.float16 if self._device == "mps" else torch.bfloat16
+
+        # specify the path to the model
+        self._vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(
+            self.model_path
+        )
+        self._tokenizer = self._vl_chat_processor.tokenizer
+
+        vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(
+            self.model_path, trust_remote_code=True, device_map=self._device
+        )
+        self._model = vl_gpt.to(self._type).eval()
+
+    @staticmethod
+    def _message_content_to_deepseek(content) -> Tuple[str, List[str]]:
+        def _ensure_url(_url):
+            if _url.startswith("data:"):
+                logging.info("Parse url by base64 decoder.")
+                # https://platform.openai.com/docs/guides/vision/uploading-base-64-encoded-images
+                # e.g. f"data:image/jpeg;base64,{base64_image}"
+                _type, data = _url.split(";")
+                _, ext = _type.split("/")
+                data = data[len("base64,") :]
+                data = base64.b64decode(data.encode("utf-8"))
+
+                with tempfile.NamedTemporaryFile(suffix=f".{ext}", delete=False) as f:
+                    f.write(data)
+                logging.info("Dump base64 data to %s", f.name)
+                return f.name
+            else:
+                if len(_url) > 2048:
+                    raise Exception(f"Image url is too long, {len(_url)} > 2048.")
+
+                return _url
+
+        def _download(_images):
+            local_images = []
+
+            # To make requests.get works
+            headers = {
+                "User-Agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/123.0.0.0 Safari/537.36"
+            }
+            with ThreadPoolExecutor() as executor:
+                for url in images:
+                    try:
+                        if os.path.exists(url):
+                            local_images.append(url)
+                            continue
+                    except Exception as e:
+                        logger.debug("Image is remote: %s, e: %s", url, e)
+                        pass
+                    # Append a placeholder
+                    local_images.append(None)
+
+                    def _fill_placeholder(_url, _index):
+                        response = requests.get(url, headers=headers)
+                        local_images[_index] = BytesIO(response.content)
+
+                    executor.submit(_fill_placeholder, url, len(local_images) - 1)
+            return local_images
+
+        if not isinstance(content, str):
+            # TODO(codingl2k1): Optimize _ensure_url
+
+            images = []
+            new_content = []
+            for c in content:
+                c_type = c.get("type")
+                if c_type == "image_url":
+                    images.append(_ensure_url(c["image_url"]["url"]))
+                elif c_type == "text":
+                    new_content.append(c["text"])
+            if images:
+                new_content.insert(0, "<image_placeholder>")
+                images = _download(images)
+            return "".join(new_content), images
+        return content, []
+
+    def chat(
+        self,
+        prompt: Union[str, List[Dict]],
+        system_prompt: Optional[str] = None,
+        chat_history: Optional[List[ChatCompletionMessage]] = None,
+        generate_config: Optional[PytorchGenerateConfig] = None,
+    ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
+        if generate_config and generate_config.get("stream"):
+            raise Exception(
+                f"Chat with model {self.model_family.model_name} does not support stream."
+            )
+        prompt, images = self._message_content_to_deepseek(prompt)
+        prompt_messages: List[Dict[str, Any]] = [
+            {
+                "role": "User",
+                "content": prompt,
+            },
+            {"role": "Assistant", "content": ""},
+        ]
+        if images:
+            prompt_messages[0]["images"] = images
+
+        # Convert openai history to qwen vl history
+        deepseek_history = []
+        for h in chat_history or []:
+            role = h["role"]
+            if role == "user":
+                content, images = self._message_content_to_deepseek(h["content"])
+                msg: Dict[str, Any] = {
+                    "role": "User",
+                    "content": content,
+                }
+                if images:
+                    msg["images"] = images
+                deepseek_history.append(msg)
+            elif role == "assistant":
+                deepseek_history.append({"role": "Assistant", "content": h["content"]})
+            else:
+                logger.error("Unexpected msg in chat history: %s", h)
+
+        deepseek_history.extend(prompt_messages)
+
+        from ....thirdparty.deepseek_vl.utils.io import load_pil_images
+
+        # load images and prepare for inputs
+        pil_images = load_pil_images(deepseek_history)
+        prepare_inputs = self._vl_chat_processor(
+            conversations=deepseek_history, images=pil_images, force_batchify=True
+        ).to(self._model.device, self._model.dtype)
+
+        # run image encoder to get the image embeddings
+        inputs_embeds = self._model.prepare_inputs_embeds(**prepare_inputs)
+
+        # run the model to get the response
+        outputs = self._model.language_model.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=prepare_inputs.attention_mask,
+            pad_token_id=self._tokenizer.eos_token_id,
+            bos_token_id=self._tokenizer.bos_token_id,
+            eos_token_id=self._tokenizer.eos_token_id,
+            max_new_tokens=512,
+            do_sample=True,
+            top_p=0.95,
+            temperature=0.2,
+            repetition_penalty=1.1,
+            use_cache=True,
+        )
+
+        answer = self._tokenizer.decode(
+            outputs[0].cpu().tolist(), skip_special_tokens=True
+        )
+
+        return ChatCompletion(
+            id="chat" + str(uuid.uuid1()),
+            object="chat.completion",
+            created=int(time.time()),
+            model=self.model_uid,
+            choices=[
+                ChatCompletionChoice(
+                    index=0,
+                    message={"role": "assistant", "content": answer},
+                    finish_reason="stop",
+                )
+            ],
+            usage=CompletionUsage(
+                prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
+            ),
+        )

xinference/model/llm/pytorch/omnilmm.py

@@ -0,0 +1,153 @@
+# Copyright 2022-2023 XProbe Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 json
+import logging
+import operator
+import tempfile
+import time
+import uuid
+from typing import Dict, Iterator, List, Optional, Tuple, Union
+
+from ....thirdparty.omnilmm.chat import OmniLMMChat, img2base64
+from ....types import (
+    ChatCompletion,
+    ChatCompletionChoice,
+    ChatCompletionChunk,
+    ChatCompletionMessage,
+    CompletionUsage,
+)
+from ...utils import select_device
+from ..llm_family import LLMFamilyV1, LLMSpecV1
+from .core import PytorchChatModel, PytorchGenerateConfig
+
+logger = logging.getLogger(__name__)
+
+
+class OmniLMMModel(PytorchChatModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._model = None
+
+    @classmethod
+    def match(
+        cls, model_family: "LLMFamilyV1", model_spec: "LLMSpecV1", quantization: str
+    ) -> bool:
+        if "OmniLMM" in model_family.model_name:
+            return True
+        return False
+
+    def load(self):
+        device = self._pytorch_model_config.get("device", "auto")
+        device = select_device(device)
+
+        self._model = OmniLMMChat(self.model_path, device_map=device)
+
+    def _message_content_to_OmniLMM(
+        self, content
+    ) -> Tuple[List[Dict[str, str]], List[Dict[str, str]]]:
+        def _ensure_url(_url):
+            if _url.startswith("data:"):
+                logging.info("Parse url by base64 decoder.")
+                # https://platform.openai.com/docs/guides/vision/uploading-base-64-encoded-images
+                # e.g. f"data:image/jpeg;base64,{base64_image}"
+                _type, data = _url.split(";")
+                _, ext = _type.split("/")
+                data = data[len("base64,") :]
+                data = base64.b64decode(data.encode("utf-8"))
+
+                with tempfile.NamedTemporaryFile(suffix=f".{ext}", delete=False) as f:
+                    f.write(data)
+                logging.info("Dump base64 data to %s", f.name)
+                return f.name
+            else:
+                if len(_url) > 2048:
+                    raise Exception(f"Image url is too long, {len(_url)} > 2048.")
+                return _url
+
+        if not isinstance(content, str):
+            images = []
+            other_content = []
+
+            for c in content:
+                if c.get("type") == "image_url":
+                    images.append(
+                        {"image": _ensure_url(c["image_url"]["url"]), "type": "image"}
+                    )
+                else:
+                    other_content.append(c)
+
+            images = sorted(images, key=operator.itemgetter("type"))
+            other_content = sorted(other_content, key=operator.itemgetter("type"))
+
+            return images, other_content
+        return [], [{"type": "text", "text": content}]
+
+    def chat(
+        self,
+        prompt: Union[str, List[Dict]],
+        system_prompt: Optional[str] = None,
+        chat_history: Optional[List[ChatCompletionMessage]] = None,
+        generate_config: Optional[PytorchGenerateConfig] = None,
+    ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
+        if generate_config and generate_config.get("stream"):
+            raise Exception(
+                f"Chat with model {self.model_family.model_name} does not support stream."
+            )
+        image_first, prompt = self._message_content_to_OmniLMM(prompt)
+
+        msgs = []
+        query_to_response: List[Dict] = []
+        image_another = []
+        for h in chat_history or []:
+            role = h["role"]
+            image_tmp, content = self._message_content_to_OmniLMM(h["content"])
+            if image_tmp != []:
+                image_another = image_tmp
+            if len(query_to_response) == 0 and role == "user":
+                query_to_response.append(
+                    {"role": "user", "content": content[0]["text"]}
+                )
+            if len(query_to_response) == 1 and role == "assistant":
+                query_to_response.append(
+                    {"role": "assistant", "content": content[0]["text"]}
+                )
+            if len(query_to_response) == 2:
+                msgs.extend(query_to_response)
+                query_to_response = []
+        if image_first != []:
+            image = image_first
+        if image_another != []:
+            image = image_another
+        im_64 = img2base64(image[0]["image"])
+        msgs.append({"role": "user", "content": prompt[0]["text"]})
+        input = {"image": im_64, "question": json.dumps(msgs, ensure_ascii=True)}
+        answer = self._model.chat(input=input)
+
+        return ChatCompletion(
+            id="chat" + str(uuid.uuid1()),
+            object="chat.completion",
+            created=int(time.time()),
+            model=self.model_uid,
+            choices=[
+                ChatCompletionChoice(
+                    index=0,
+                    message={"role": "assistant", "content": answer},
+                    finish_reason="stop",
+                )
+            ],
+            usage=CompletionUsage(
+                prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
+            ),
+        )

xinference/model/llm/utils.py

@@ -421,6 +421,16 @@ Begin!"""
                 else:
                     ret += f"{role}".rstrip()
             return ret
+        elif prompt_style.style_name == "MINICPM-2B":
+            ret = ""
+            for message in chat_history:
+                content = message["content"] or ""
+                role = get_role(message["role"])
+                if role == "user":
+                    ret += "<用户>" + content.strip()
+                else:
+                    ret += "<AI>" + content.strip()
+            return ret
         else:
             raise ValueError(f"Invalid prompt style: {prompt_style.style_name}")
 
@@ -585,7 +595,7 @@ Begin!"""
     def _tool_calls_completion(cls, model_family, model_uid, c, tools):
         _id = str(uuid.uuid4())
         family = model_family.model_family or model_family.model_name
-        if "gorilla-openfunctions-v1" == family:
+        if family in ["gorilla-openfunctions-v1", "gorilla-openfunctions-v2"]:
             content, func, args = cls._eval_gorilla_openfunctions_arguments(c, tools)
         elif "chatglm3" == family:
             content, func, args = cls._eval_chatglm3_arguments(c, tools)

xinference/model/llm/vllm/core.py

@@ -86,6 +86,7 @@ VLLM_SUPPORTED_CHAT_MODELS = [
     "vicuna-v1.3",
     "vicuna-v1.5",
     "baichuan-chat",
+    "baichuan-2-chat",
     "internlm-chat-7b",
     "internlm-chat-8k",
     "internlm-chat-20b",
@@ -99,6 +100,8 @@ VLLM_SUPPORTED_CHAT_MODELS = [
     "mistral-instruct-v0.2",
     "mixtral-instruct-v0.1",
     "chatglm3",
+    "deepseek-chat",
+    "deepseek-coder-instruct",
 ]
 if VLLM_INSTALLED and vllm.__version__ >= "0.3.0":
     VLLM_SUPPORTED_CHAT_MODELS.append("qwen1.5-chat")

xinference/thirdparty/deepseek_vl/__init__.py

@@ -0,0 +1,31 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+
+# check if python version is above 3.10
+import sys
+
+if sys.version_info >= (3, 10):
+    print("Python version is above 3.10, patching the collections module.")
+    # Monkey patch collections
+    import collections
+    import collections.abc
+
+    for type_name in collections.abc.__all__:
+        setattr(collections, type_name, getattr(collections.abc, type_name))

xinference/thirdparty/deepseek_vl/models/__init__.py

@@ -0,0 +1,28 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from .image_processing_vlm import VLMImageProcessor
+from .modeling_vlm import MultiModalityCausalLM
+from .processing_vlm import VLChatProcessor
+
+__all__ = [
+    "VLMImageProcessor",
+    "VLChatProcessor",
+    "MultiModalityCausalLM",
+]