File size: 7,774 Bytes
d8b5140
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211

"""
    该文件中主要包含2个函数

    不具备多线程能力的函数:
    1. predict: 正常对话时使用,具备完备的交互功能,不可多线程

    具备多线程调用能力的函数
    2. predict_no_ui_long_connection:在实验过程中发现调用predict_no_ui处理长文档时,和openai的连接容易断掉,这个函数用stream的方式解决这个问题,同样支持多线程
"""
import tiktoken
from functools import wraps, lru_cache
from concurrent.futures import ThreadPoolExecutor

from .bridge_chatgpt import predict_no_ui_long_connection as chatgpt_noui
from .bridge_chatgpt import predict as chatgpt_ui

from .bridge_chatglm import predict_no_ui_long_connection as chatglm_noui
from .bridge_chatglm import predict as chatglm_ui

# from .bridge_tgui import predict_no_ui_long_connection as tgui_noui
# from .bridge_tgui import predict as tgui_ui

colors = ['#FF00FF', '#00FFFF', '#FF0000', '#990099', '#009999', '#990044']

class LazyloadTiktoken(object):
    def __init__(self, model):
        self.model = model

    @staticmethod
    @lru_cache(maxsize=128)
    def get_encoder(model):
        print('正在加载tokenizer,如果是第一次运行,可能需要一点时间下载参数')
        tmp = tiktoken.encoding_for_model(model)
        print('加载tokenizer完毕')
        return tmp
    
    def encode(self, *args, **kwargs):
        encoder = self.get_encoder(self.model) 
        return encoder.encode(*args, **kwargs)
    
    def decode(self, *args, **kwargs):
        encoder = self.get_encoder(self.model) 
        return encoder.decode(*args, **kwargs)
    
tokenizer_gpt35 = LazyloadTiktoken("gpt-3.5-turbo")
tokenizer_gpt4 = LazyloadTiktoken("gpt-4")
get_token_num_gpt35 = lambda txt: len(tokenizer_gpt35.encode(txt, disallowed_special=()))
get_token_num_gpt4 = lambda txt: len(tokenizer_gpt4.encode(txt, disallowed_special=()))

model_info = {
    # openai
    "gpt-3.5-turbo": {
        "fn_with_ui": chatgpt_ui,
        "fn_without_ui": chatgpt_noui,
        "endpoint": "https://api.openai.com/v1/chat/completions",
        "max_token": 4096,
        "tokenizer": tokenizer_gpt35,
        "token_cnt": get_token_num_gpt35,
    },

    "gpt-4": {
        "fn_with_ui": chatgpt_ui,
        "fn_without_ui": chatgpt_noui,
        "endpoint": "https://api.openai.com/v1/chat/completions",
        "max_token": 8192,
        "tokenizer": tokenizer_gpt4,
        "token_cnt": get_token_num_gpt4,
    },

    # api_2d
    "api2d-gpt-3.5-turbo": {
        "fn_with_ui": chatgpt_ui,
        "fn_without_ui": chatgpt_noui,
        "endpoint": "https://openai.api2d.net/v1/chat/completions",
        "max_token": 4096,
        "tokenizer": tokenizer_gpt35,
        "token_cnt": get_token_num_gpt35,
    },

    "api2d-gpt-4": {
        "fn_with_ui": chatgpt_ui,
        "fn_without_ui": chatgpt_noui,
        "endpoint": "https://openai.api2d.net/v1/chat/completions",
        "max_token": 8192,
        "tokenizer": tokenizer_gpt4,
        "token_cnt": get_token_num_gpt4,
    },

    # chatglm
    "chatglm": {
        "fn_with_ui": chatglm_ui,
        "fn_without_ui": chatglm_noui,
        "endpoint": None,
        "max_token": 1024,
        "tokenizer": tokenizer_gpt35,
        "token_cnt": get_token_num_gpt35,
    },

}


def LLM_CATCH_EXCEPTION(f):
    """
    装饰器函数,将错误显示出来
    """
    def decorated(inputs, llm_kwargs, history, sys_prompt, observe_window, console_slience):
        try:
            return f(inputs, llm_kwargs, history, sys_prompt, observe_window, console_slience)
        except Exception as e:
            from toolbox import get_conf
            import traceback
            proxies, = get_conf('proxies')
            tb_str = '\n```\n' + traceback.format_exc() + '\n```\n'
            observe_window[0] = tb_str
            return tb_str
    return decorated


def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, observe_window, console_slience=False):
    """
    发送至LLM,等待回复,一次性完成,不显示中间过程。但内部用stream的方法避免中途网线被掐。
    inputs:
        是本次问询的输入
    sys_prompt:
        系统静默prompt
    llm_kwargs:
        LLM的内部调优参数
    history:
        是之前的对话列表
    observe_window = None:
        用于负责跨越线程传递已经输出的部分,大部分时候仅仅为了fancy的视觉效果,留空即可。observe_window[0]:观测窗。observe_window[1]:看门狗
    """
    import threading, time, copy

    model = llm_kwargs['llm_model']
    n_model = 1
    if '&' not in model:
        assert not model.startswith("tgui"), "TGUI不支持函数插件的实现"

        # 如果只询问1个大语言模型:
        method = model_info[model]["fn_without_ui"]
        return method(inputs, llm_kwargs, history, sys_prompt, observe_window, console_slience)
    else:
        # 如果同时询问多个大语言模型:
        executor = ThreadPoolExecutor(max_workers=4)
        models = model.split('&')
        n_model = len(models)
        
        window_len = len(observe_window)
        assert window_len==3
        window_mutex = [["", time.time(), ""] for _ in range(n_model)] + [True]

        futures = []
        for i in range(n_model):
            model = models[i]
            method = model_info[model]["fn_without_ui"]
            llm_kwargs_feedin = copy.deepcopy(llm_kwargs)
            llm_kwargs_feedin['llm_model'] = model
            future = executor.submit(LLM_CATCH_EXCEPTION(method), inputs, llm_kwargs_feedin, history, sys_prompt, window_mutex[i], console_slience)
            futures.append(future)

        def mutex_manager(window_mutex, observe_window):
            while True:
                time.sleep(0.5)
                if not window_mutex[-1]: break
                # 看门狗(watchdog)
                for i in range(n_model): 
                    window_mutex[i][1] = observe_window[1]
                # 观察窗(window)
                chat_string = []
                for i in range(n_model):
                    chat_string.append( f"【{str(models[i])} 说】: <font color=\"{colors[i]}\"> {window_mutex[i][0]} </font>" )
                res = '<br/><br/>\n\n---\n\n'.join(chat_string)
                # # # # # # # # # # #
                observe_window[0] = res

        t_model = threading.Thread(target=mutex_manager, args=(window_mutex, observe_window), daemon=True)
        t_model.start()

        return_string_collect = []
        while True:
            worker_done = [h.done() for h in futures]
            if all(worker_done):
                executor.shutdown()
                break
            time.sleep(1)

        for i, future in enumerate(futures):  # wait and get
            return_string_collect.append( f"【{str(models[i])} 说】: <font color=\"{colors[i]}\"> {future.result()} </font>" )

        window_mutex[-1] = False # stop mutex thread
        res = '<br/>\n\n---\n\n'.join(return_string_collect)
        return res


def predict(inputs, llm_kwargs, *args, **kwargs):
    """
    发送至LLM,流式获取输出。
    用于基础的对话功能。
    inputs 是本次问询的输入
    top_p, temperature是LLM的内部调优参数
    history 是之前的对话列表(注意无论是inputs还是history,内容太长了都会触发token数量溢出的错误)
    chatbot 为WebUI中显示的对话列表,修改它,然后yeild出去,可以直接修改对话界面内容
    additional_fn代表点击的哪个按钮,按钮见functional.py
    """

    method = model_info[llm_kwargs['llm_model']]["fn_with_ui"]
    yield from method(inputs, llm_kwargs, *args, **kwargs)