CC/gpt_academic
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

ChatGLM改成多进程运行

Browse Source
This commit is contained in:
qingxu fu
2023-04-15 19:09:03 +08:00
parent 294ac338bd
commit 6aba339538
7 changed files with 270 additions and 212 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
request_llm/bridge_all.py
View File

@@ -31,6 +31,24 @@ methods = {
"tgui-ui": tgui_ui,
}
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
colors = ['#FF00FF', '#00FFFF', '#FF0000''#990099', '#009999', '#990044']
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。
@@ -62,17 +80,13 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, obser
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=16)
executor = ThreadPoolExecutor(max_workers=4)
models = model.split('&')
n_model = len(models)
window_len = len(observe_window)
if window_len==0:
window_mutex = [[] for _ in range(n_model)] + [True]
elif window_len==1:
window_mutex = [[""] for _ in range(n_model)] + [True]
elif window_len==2:
window_mutex = [["", time.time()] for _ in range(n_model)] + [True]
assert window_len==3
window_mutex = [["", time.time(), ""] for _ in range(n_model)] + [True]
futures = []
for i in range(n_model):
@@ -85,12 +99,12 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, obser
method = methods['tgui-no-ui']
llm_kwargs_feedin = copy.deepcopy(llm_kwargs)
llm_kwargs_feedin['llm_model'] = model
future = executor.submit(method, inputs, llm_kwargs_feedin, history, sys_prompt, window_mutex[i], console_slience)
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(0.2)
time.sleep(0.5)
if not window_mutex[-1]: break
# 看门狗（watchdog）
for i in range(n_model):
@@ -98,8 +112,8 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, obser
# 观察窗（window）
chat_string = []
for i in range(n_model):
chat_string.append( f"[{str(models[i])} 说]: {window_mutex[i][0]}" )
res = '\n\n---\n\n'.join(chat_string)
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
@@ -107,10 +121,18 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history, sys_prompt, obser
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]: {future.result()}" )
return_string_collect.append( f"【{str(models[i])} 说】: <font color=\"{colors[i]}\"> {future.result()} </font>" )
window_mutex[-1] = False # stop mutex thread
res = '\n\n---\n\n'.join(return_string_collect)
res = '<br/>\n\n---\n\n'.join(return_string_collect)
return res

98
request_llm/bridge_chatglm.py
View File

@@ -3,35 +3,69 @@ from transformers import AutoModel, AutoTokenizer
import time
import importlib
from toolbox import update_ui, get_conf
from multiprocessing import Process, Pipe
#################################################################################
class GetGLMHandle(Process):
def __init__(self):
super().__init__(daemon=True)
self.parent, self.child = Pipe()
self.chatglm_model = None
self.chatglm_tokenizer = None
self.start()
print('初始化')
def ready(self):
return self.chatglm_model is not None
global chatglm_model, chatglm_tokenizer
chatglm_model = None
chatglm_tokenizer = None
def model_loader():
global chatglm_model, chatglm_tokenizer
if chatglm_tokenizer is None:
chatglm_tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
if chatglm_model is None: # 尚未加载
device, = get_conf('LOCAL_MODEL_DEVICE')
if device=='cpu':
chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).float()
else:
chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
chatglm_model = chatglm_model.eval()
chatglm_model = chatglm_model.eval()
def run(self):
while True:
try:
if self.chatglm_model is None:
self.chatglm_tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
device, = get_conf('LOCAL_MODEL_DEVICE')
if device=='cpu':
self.chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).float()
else:
self.chatglm_model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
self.chatglm_model = self.chatglm_model.eval()
break
else:
break
except:
pass
while True:
kwargs = self.child.recv()
try:
for response, history in self.chatglm_model.stream_chat(self.chatglm_tokenizer, **kwargs):
self.child.send(response)
except:
self.child.send('[Local Message] Call ChatGLM fail.')
self.child.send('[Finish]')
def stream_chat(self, **kwargs):
self.parent.send(kwargs)
while True:
res = self.parent.recv()
if res != '[Finish]':
yield res
else:
break
return
global glm_handle
glm_handle = None
#################################################################################
def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="", observe_window=None, console_slience=False):
"""
多线程方法
函数的说明请见 request_llm/bridge_all.py
"""
global chatglm_model, chatglm_tokenizer
if chatglm_model is None:
observe_window[0] = "ChatGLM尚未加载，加载需要一段时间 ……"
global glm_handle
if glm_handle is None:
glm_handle = GetGLMHandle()
observe_window[0] = "ChatGLM尚未加载，加载需要一段时间。注意，取决于`config.py`的配置，ChatGLM消耗大量的内存（CPU）或显存（GPU），也许会导致低配计算机卡死 ……"
model_loader()
# chatglm 没有 sys_prompt 接口，因此把prompt加入 history
history_feedin = []
for i in range(len(history)//2):
@@ -40,29 +74,27 @@ def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="",
watch_dog_patience = 5 # 看门狗 (watchdog) 的耐心, 设置5秒即可
response = ""
for response, history in chatglm_model.stream_chat(chatglm_tokenizer, inputs, history=history_feedin, max_length=llm_kwargs['max_length'],
top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
# 观测窗，把已经获取的数据显示出去
for response in glm_handle.stream_chat(query=inputs, history=history_feedin, max_length=llm_kwargs['max_length'], top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
observe_window[0] = response
# 看门狗 (watchdog)，如果超过期限没有喂狗，则终止
if len(observe_window) >= 2:
if (time.time()-observe_window[1]) > watch_dog_patience:
raise RuntimeError("程序终止。")
# if not console_slience:
# print(response)
return response
def predict(inputs, llm_kwargs, plugin_kwargs, chatbot, history=[], system_prompt='', stream = True, additional_fn=None):
"""
单线程方法
函数的说明请见 request_llm/bridge_all.py
"""
global chatglm_model, chatglm_tokenizer
chatbot.append((inputs, ""))
if chatglm_model is None:
chatbot[-1] = (inputs, "ChatGLM尚未加载，加载需要一段时间 ……")
global glm_handle
if glm_handle is None:
glm_handle = GetGLMHandle()
chatbot[-1] = (inputs, "ChatGLM尚未加载，加载需要一段时间。注意，取决于`config.py`的配置，ChatGLM消耗大量的内存（CPU）或显存（GPU），也许会导致低配计算机卡死 ……")
yield from update_ui(chatbot=chatbot, history=[])
model_loader()
if additional_fn is not None:
import core_functional
@@ -71,13 +103,11 @@ def predict(inputs, llm_kwargs, plugin_kwargs, chatbot, history=[], system_promp
if "PreProcess" in core_functional[additional_fn]: inputs = core_functional[additional_fn]["PreProcess"](inputs) # 获取预处理函数（如果有的话）
inputs = core_functional[additional_fn]["Prefix"] + inputs + core_functional[additional_fn]["Suffix"]
history_feedin = []
for i in range(len(history)//2):
history_feedin.append(["What can I do?", system_prompt] )
history_feedin.append([history[2*i], history[2*i+1]] )
for response, history in chatglm_model.stream_chat(chatglm_tokenizer, inputs, history=history_feedin, max_length=llm_kwargs['max_length'],
top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
for response in glm_handle.stream_chat(query=inputs, history=history_feedin, max_length=llm_kwargs['max_length'], top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
chatbot[-1] = (inputs, response)
yield from update_ui(chatbot=chatbot, history=history)