xinference 0.11.2.post1__py3-none-any.whl → 0.12.0__py3-none-any.whl

xinference/model/llm/pytorch/baichuan.py

@@ -73,7 +73,8 @@ class BaichuanPytorchChatModel(PytorchChatModel):
     ) -> bool:
         if llm_spec.model_format != "pytorch":
             return False
-        if llm_family.model_name not in ["baichuan-chat", "baichuan-2-chat"]:
+        model_family = llm_family.model_family or llm_family.model_name
+        if model_family not in ["baichuan-chat", "baichuan-2-chat"]:
             return False
         if "chat" not in llm_family.model_ability:
             return False

xinference/model/llm/pytorch/chatglm.py

@@ -82,7 +82,8 @@ class ChatglmPytorchChatModel(PytorchChatModel):
     ) -> bool:
         if llm_spec.model_format != "pytorch":
             return False
-        if "chatglm" not in llm_family.model_name:
+        model_family = llm_family.model_family or llm_family.model_name
+        if "chatglm" not in model_family and "glm4" not in model_family:
             return False
         if "chat" not in llm_family.model_ability:
             return False

xinference/model/llm/pytorch/cogvlm2.py

@@ -0,0 +1,316 @@
+# 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 time
+import uuid
+from concurrent.futures import ThreadPoolExecutor
+from io import BytesIO
+from typing import Dict, Iterator, List, Optional, Tuple, Union
+
+import requests
+import torch
+from PIL import Image
+
+from ....model.utils import select_device
+from ....types import (
+    ChatCompletion,
+    ChatCompletionChunk,
+    ChatCompletionMessage,
+    Completion,
+    CompletionChoice,
+    CompletionChunk,
+    CompletionUsage,
+)
+from ..llm_family import LLMFamilyV1, LLMSpecV1
+from .core import PytorchChatModel, PytorchGenerateConfig
+
+logger = logging.getLogger(__name__)
+
+IMAGENET_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_STD = (0.229, 0.224, 0.225)
+
+
+class CogVLM2Model(PytorchChatModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._torch_type = None
+        self._device = None
+        self._tokenizer = None
+        self._model = None
+
+    @classmethod
+    def match(
+        cls, model_family: "LLMFamilyV1", model_spec: "LLMSpecV1", quantization: str
+    ) -> bool:
+        family = model_family.model_family or model_family.model_name
+        if "cogvlm" in family.lower():
+            return True
+        return False
+
+    def load(self, **kwargs):
+        from transformers import AutoModelForCausalLM, AutoTokenizer
+        from transformers.generation import GenerationConfig
+
+        device = self._pytorch_model_config.get("device", "auto")
+        self._device = select_device(device)
+        self._torch_type = (
+            torch.bfloat16
+            if torch.cuda.is_available() and torch.cuda.get_device_capability()[0] >= 8
+            else torch.float16
+        )
+
+        self._tokenizer = AutoTokenizer.from_pretrained(
+            self.model_path,
+            trust_remote_code=True,
+        )
+
+        self._model = AutoModelForCausalLM.from_pretrained(
+            self.model_path,
+            torch_dtype=self._torch_type,
+            trust_remote_code=True,
+            low_cpu_mem_usage=True,
+            device_map="auto",
+        ).eval()
+
+        # Specify hyperparameters for generation
+        self._model.generation_config = GenerationConfig.from_pretrained(
+            self.model_path,
+            trust_remote_code=True,
+        )
+
+    def _message_content_to_cogvlm2(self, content):
+        def _load_image(_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"))
+                return Image.open(BytesIO(data)).convert("RGB")
+            else:
+                try:
+                    response = requests.get(_url)
+                except requests.exceptions.MissingSchema:
+                    return Image.open(_url).convert("RGB")
+                else:
+                    return Image.open(BytesIO(response.content)).convert("RGB")
+
+        if not isinstance(content, str):
+            texts = []
+            image_urls = []
+            for c in content:
+                c_type = c.get("type")
+                if c_type == "text":
+                    texts.append(c["text"])
+                elif c_type == "image_url":
+                    image_urls.append(c["image_url"]["url"])
+            image_futures = []
+            with ThreadPoolExecutor() as executor:
+                for image_url in image_urls:
+                    fut = executor.submit(_load_image, image_url)
+                    image_futures.append(fut)
+            images = [fut.result() for fut in image_futures]
+            text = " ".join(texts)
+            if len(images) == 0:
+                return text, None
+            elif len(images) == 1:
+                return text, images
+            else:
+                raise RuntimeError(
+                    "Only one image per message is supported by CogVLM2."
+                )
+        return content, None
+
+    def _history_content_to_cogvlm2(
+        self, system_prompt: str, chat_history: List[ChatCompletionMessage]
+    ):
+        def _image_to_piexl_values(image):
+            if image.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 = image.split(";")
+                _, ext = _type.split("/")
+                data = data[len("base64,") :]
+                data = base64.b64decode(data.encode("utf-8"))
+                return Image.open(BytesIO(data)).convert("RGB")
+            else:
+                try:
+                    response = requests.get(image)
+                except requests.exceptions.MissingSchema:
+                    return Image.open(image).convert("RGB")
+                else:
+                    return Image.open(BytesIO(response.content)).convert("RGB")
+
+        query = system_prompt
+        history: List[Tuple] = []
+        pixel_values = None
+        for i in range(0, len(chat_history), 2):
+            user = chat_history[i]["content"]
+            if isinstance(user, List):
+                for content in user:
+                    c_type = content.get("type")
+                    if c_type == "text":
+                        user = content["text"]
+                    elif c_type == "image_url" and not pixel_values:
+                        pixel_values = _image_to_piexl_values(
+                            content["image_url"]["url"]
+                        )
+            assistant = chat_history[i + 1]["content"]
+            query = query + f" USER: {user} ASSISTANT:"
+            history.append((query, assistant))
+            query = query + f" {assistant}"
+        return query, history, [pixel_values]
+
+    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]]:
+        system_prompt = system_prompt if system_prompt else ""
+        stream = generate_config.get("stream", False) if generate_config else False
+
+        sanitized_config = {
+            "pad_token_id": 128002,
+            "max_new_tokens": generate_config.get("max_tokens", 512)
+            if generate_config
+            else 512,
+        }
+
+        content, image = self._message_content_to_cogvlm2(prompt)
+
+        history = []
+        query = ""
+        history_image = None
+        if chat_history:
+            query, history, history_image = self._history_content_to_cogvlm2(
+                system_prompt, chat_history
+            )
+
+        if image and history_image:
+            history = []
+            query = system_prompt + f" USER: {content} ASSISTANT:"
+        else:
+            image = image if image else history_image
+            query = query + f" USER: {content} ASSISTANT:"
+
+        input_by_model = self._model.build_conversation_input_ids(
+            self._tokenizer,
+            query=query,
+            history=history,
+            images=image,
+            template_version="chat",
+        )
+
+        inputs = {
+            "input_ids": input_by_model["input_ids"].unsqueeze(0).to(self._device),
+            "token_type_ids": input_by_model["token_type_ids"]
+            .unsqueeze(0)
+            .to(self._device),
+            "attention_mask": input_by_model["attention_mask"]
+            .unsqueeze(0)
+            .to(self._device),
+            "images": [
+                [input_by_model["images"][0].to(self._device).to(self._torch_type)]
+            ]
+            if image is not None
+            else None,
+        }
+
+        if stream:
+            it = self._streaming_chat_response(inputs, sanitized_config)
+            return self._to_chat_completion_chunks(it)
+        else:
+            with torch.no_grad():
+                outputs = self._model.generate(**inputs, **sanitized_config)
+                outputs = outputs[:, inputs["input_ids"].shape[1] :]
+                response = self._tokenizer.decode(outputs[0])
+                response = response.split("<|end_of_text|>")[0]
+
+            chunk = Completion(
+                id=str(uuid.uuid1()),
+                object="text_completion",
+                created=int(time.time()),
+                model=self.model_uid,
+                choices=[
+                    CompletionChoice(
+                        index=0, text=response, finish_reason="stop", logprobs=None
+                    )
+                ],
+                usage=CompletionUsage(
+                    prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
+                ),
+            )
+            return self._to_chat_completion(chunk)
+
+    def _streaming_chat_response(
+        self, inputs: Dict, config: Dict
+    ) -> Iterator[CompletionChunk]:
+        from threading import Thread
+
+        from transformers import TextIteratorStreamer
+
+        streamer = TextIteratorStreamer(
+            self._tokenizer, skip_prompt=True, skip_special_tokens=True
+        )
+        generation_kwargs = {
+            "input_ids": inputs["input_ids"],
+            "attention_mask": inputs["attention_mask"],
+            "token_type_ids": inputs["token_type_ids"],
+            "images": inputs["images"],
+            "max_new_tokens": config["max_new_tokens"],
+            "pad_token_id": config["pad_token_id"],
+            "streamer": streamer,
+        }
+
+        thread = Thread(target=self._model.generate, kwargs=generation_kwargs)
+        thread.start()
+
+        completion_id = str(uuid.uuid1())
+        for new_text in streamer:
+            chunk = CompletionChunk(
+                id=completion_id,
+                object="text_completion",
+                created=int(time.time()),
+                model=self.model_uid,
+                choices=[
+                    CompletionChoice(
+                        index=0, text=new_text, finish_reason=None, logprobs=None
+                    )
+                ],
+                usage=CompletionUsage(
+                    prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
+                ),
+            )
+            yield chunk
+
+        completion_choice = CompletionChoice(
+            text="", index=0, logprobs=None, finish_reason="stop"
+        )
+        chunk = CompletionChunk(
+            id=completion_id,
+            object="text_completion",
+            created=int(time.time()),
+            model=self.model_uid,
+            choices=[completion_choice],
+            usage=CompletionUsage(
+                prompt_tokens=-1, completion_tokens=-1, total_tokens=-1
+            ),
+        )
+        yield chunk

xinference/model/llm/pytorch/core.py

@@ -17,6 +17,7 @@ import logging
 import os
 from typing import Iterable, Iterator, List, Optional, Union
 
+from ....core.scheduler import InferenceRequest
 from ....device_utils import (
     get_device_preferred_dtype,
     gpu_count,
@@ -40,6 +41,7 @@ from ...utils import select_device
 from ..core import LLM
 from ..llm_family import LLMFamilyV1, LLMSpecV1
 from ..utils import ChatModelMixin
+from .utils import get_context_length, get_max_src_len
 
 logger = logging.getLogger(__name__)
 
@@ -53,6 +55,11 @@ NON_DEFAULT_MODEL_LIST: List[str] = [
     "chatglm2",
     "chatglm2-32k",
     "chatglm2-128k",
+    "chatglm3",
+    "chatglm3-32k",
+    "chatglm3-128k",
+    "glm4-chat",
+    "glm4-chat-1m",
     "llama-2",
     "llama-2-chat",
     "internlm2-chat",
@@ -62,6 +69,9 @@ NON_DEFAULT_MODEL_LIST: List[str] = [
     "deepseek-vl-chat",
     "internvl-chat",
     "mini-internvl-chat",
+    "cogvlm2",
+    "MiniCPM-Llama3-V-2_5",
+    "glm-4v",
 ]
 
 
@@ -95,6 +105,7 @@ class PytorchModel(LLM):
         pytorch_model_config.setdefault("gptq_act_order", False)
         pytorch_model_config.setdefault("device", "auto")
         pytorch_model_config.setdefault("trust_remote_code", True)
+        pytorch_model_config.setdefault("max_num_seqs", 16)
         return pytorch_model_config
 
     def _sanitize_generate_config(
@@ -453,6 +464,7 @@ class PytorchChatModel(PytorchModel, ChatModelMixin):
             pytorch_model_config,
             peft_model,
         )
+        self._context_len = None
 
     def _sanitize_generate_config(
         self,
@@ -496,13 +508,8 @@ class PytorchChatModel(PytorchModel, ChatModelMixin):
         chat_history: Optional[List[ChatCompletionMessage]] = None,
         generate_config: Optional[PytorchGenerateConfig] = None,
     ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
-        assert self.model_family.prompt_style is not None
-        prompt_style = self.model_family.prompt_style.copy()
-        if system_prompt:
-            prompt_style.system_prompt = system_prompt
-        chat_history = chat_history or []
         tools = generate_config.pop("tools", []) if generate_config else None
-        full_prompt = self.get_prompt(prompt, chat_history, prompt_style, tools=tools)
+        full_prompt = self._get_full_prompt(prompt, system_prompt, chat_history, tools)
 
         generate_config = self._sanitize_generate_config(generate_config)
         # TODO(codingl2k1): qwen hacky to set stop for function call.
@@ -530,3 +537,82 @@ class PytorchChatModel(PytorchModel, ChatModelMixin):
                     self.model_family, self.model_uid, c, tools
                 )
             return self._to_chat_completion(c)
+
+    def load(self):
+        super().load()
+        self._context_len = get_context_length(self._model.config)
+
+    def _get_full_prompt(self, prompt, system_prompt, chat_history, tools):
+        assert self.model_family.prompt_style is not None
+        prompt_style = self.model_family.prompt_style.copy()
+        if system_prompt:
+            prompt_style.system_prompt = system_prompt
+        chat_history = chat_history or []
+        full_prompt = ChatModelMixin.get_prompt(
+            prompt, chat_history, prompt_style, tools=tools
+        )
+        return full_prompt
+
+    def get_max_num_seqs(self) -> int:
+        return self._pytorch_model_config.get("max_num_seqs")  # type: ignore
+
+    def batch_inference(self, req_list: List[InferenceRequest]):
+        from .utils import batch_inference_one_step
+
+        for r in req_list:
+            if r.sanitized_generate_config is None:
+                r.sanitized_generate_config = self._sanitize_generate_config(
+                    r.generate_config
+                )
+            if r.is_prefill:
+                # check some generate params
+                max_src_len = get_max_src_len(self._context_len, r)  # type: ignore
+                if max_src_len < 0:
+                    r.stopped = True
+                    r.error_msg = "Max tokens exceeds model's max length"
+                    continue
+                if r.stream_interval <= 0:
+                    r.stopped = True
+                    r.error_msg = "`stream_interval` must be greater than 0"
+                    continue
+                stop_str = r.sanitized_generate_config.get("stop", None)
+                if stop_str and (
+                    not (isinstance(stop_str, str) or isinstance(stop_str, Iterable))
+                ):
+                    r.stopped = True
+                    r.error_msg = "Invalid `stop` field type"
+                    continue
+                r.full_prompt = self._get_full_prompt(
+                    r.prompt, r.system_prompt, r.chat_history, None
+                )
+
+        assert isinstance(self._context_len, int)
+        batch_inference_one_step(
+            req_list,
+            self.model_uid,
+            self._model,
+            self._tokenizer,
+            self._device,
+            self._context_len,
+        )
+        for req in req_list:
+            if req.stream and req.error_msg is None:
+                if req.completion:
+                    results = []
+                    for i, c in enumerate(req.completion):
+                        if c == "<bos_stream>":
+                            results.append(
+                                self._get_first_chat_completion_chunk(
+                                    req.completion[i + 1]
+                                )
+                            )
+                        elif c == "<eos_stream>":
+                            break
+                        else:
+                            results.append(self._to_chat_completion_chunk(c))
+
+                    if req.stopped and req.include_usage:
+                        results.append(
+                            self._get_final_chat_completion_chunk(req.completion[-1])
+                        )
+                    req.completion = results