Langchain-Chatchat

1
import json
2
import sys
3

4
from fastchat.conversation import Conversation
5
from configs import TEMPERATURE
6
from http import HTTPStatus
7
from typing import List, Literal, Dict
8

9
from fastchat import conversation as conv
10
from server.model_workers.base import *
11
from server.model_workers.base import ApiEmbeddingsParams
12
from configs import logger, log_verbose
13

14

15
class QwenWorker(ApiModelWorker):
16
    DEFAULT_EMBED_MODEL = "text-embedding-v1"
17

18
    def __init__(
19
        self,
20
        *,
21
        version: Literal["qwen-turbo", "qwen-plus"] = "qwen-turbo",
22
        model_names: List[str] = ["qwen-api"],
23
        controller_addr: str = None,
24
        worker_addr: str = None,
25
        **kwargs,
26
    ):
27
        kwargs.update(model_names=model_names, controller_addr=controller_addr, worker_addr=worker_addr)
28
        kwargs.setdefault("context_len", 16384)
29
        super().__init__(**kwargs)
30
        self.version = version
31

32
    def do_chat(self, params: ApiChatParams) -> Dict:
33
        import dashscope
34
        params.load_config(self.model_names[0])
35
        if log_verbose:
36
            logger.info(f'{self.__class__.__name__}:params: {params}')
37

38
        gen = dashscope.Generation()
39
        responses = gen.call(
40
            model=params.version,
41
            temperature=params.temperature,
42
            api_key=params.api_key,
43
            messages=params.messages,
44
            result_format='message',  # set the result is message format.
45
            stream=True,
46
        )
47

48
        for resp in responses:
49
            if resp["status_code"] == 200:
50
                if choices := resp["output"]["choices"]:
51
                    yield {
52
                        "error_code": 0,
53
                        "text": choices[0]["message"]["content"],
54
                    }
55
            else:
56
                data = {
57
                    "error_code": resp["status_code"],
58
                    "text": resp["message"],
59
                    "error": {
60
                        "message": resp["message"],
61
                        "type": "invalid_request_error",
62
                        "param": None,
63
                        "code": None,
64
                    }
65
                }
66
                self.logger.error(f"请求千问 API 时发生错误：{data}")
67
                yield data
68

69
    def do_embeddings(self, params: ApiEmbeddingsParams) -> Dict:
70
        import dashscope
71
        params.load_config(self.model_names[0])
72
        if log_verbose:
73
            logger.info(f'{self.__class__.__name__}:params: {params}')
74
        result = []
75
        i = 0
76
        while i < len(params.texts):
77
            texts = params.texts[i:i+25]
78
            resp = dashscope.TextEmbedding.call(
79
                model=params.embed_model or self.DEFAULT_EMBED_MODEL,
80
                input=texts, # 最大25行
81
                api_key=params.api_key,
82
            )
83
            if resp["status_code"] != 200:
84
                data = {
85
                            "code": resp["status_code"],
86
                            "msg": resp.message,
87
                            "error": {
88
                                "message": resp["message"],
89
                                "type": "invalid_request_error",
90
                                "param": None,
91
                                "code": None,
92
                            }
93
                        }
94
                self.logger.error(f"请求千问 API 时发生错误：{data}")
95
                return data
96
            else:
97
                embeddings = [x["embedding"] for x in resp["output"]["embeddings"]]
98
                result += embeddings
99
            i += 25
100
        return {"code": 200, "data": result}
101

102
    def get_embeddings(self, params):
103
        print("embedding")
104
        print(params)
105

106
    def make_conv_template(self, conv_template: str = None, model_path: str = None) -> Conversation:
107
        return conv.Conversation(
108
            name=self.model_names[0],
109
            system_message="你是一个聪明、对人类有帮助的人工智能，你可以对人类提出的问题给出有用、详细、礼貌的回答。",
110
            messages=[],
111
            roles=["user", "assistant", "system"],
112
            sep="\n### ",
113
            stop_str="###",
114
        )
115

116

117
if __name__ == "__main__":
118
    import uvicorn
119
    from server.utils import MakeFastAPIOffline
120
    from fastchat.serve.model_worker import app
121

122
    worker = QwenWorker(
123
        controller_addr="http://127.0.0.1:20001",
124
        worker_addr="http://127.0.0.1:20007",
125
    )
126
    sys.modules["fastchat.serve.model_worker"].worker = worker
127
    MakeFastAPIOffline(app)
128
    uvicorn.run(app, port=20007)
129