gpt-academicc / request_llm /bridge_all.py
3v324v23's picture
lazy load +decode 方法
c37c49d
raw
history blame
7.77 kB
"""
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)