better local model interaction
This commit is contained in:
binary-husky
parent
08f036aafd
commit
136162ec0d
@ -1,42 +1,29 @@
|
||||
model_name = "ChatGLM"
|
||||
cmd_to_install = "`pip install -r request_llms/requirements_chatglm.txt`"
|
||||
|
||||
|
||||
from transformers import AutoModel, AutoTokenizer
|
||||
import time
|
||||
import threading
|
||||
import importlib
|
||||
from toolbox import update_ui, get_conf, ProxyNetworkActivate
|
||||
from multiprocessing import Process, Pipe
|
||||
from toolbox import get_conf, ProxyNetworkActivate
|
||||
from .local_llm_class import LocalLLMHandle, get_local_llm_predict_fns, SingletonLocalLLM
|
||||
|
||||
load_message = "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 ……"
|
||||
|
||||
#################################################################################
|
||||
class GetGLMHandle(Process):
|
||||
def __init__(self):
|
||||
super().__init__(daemon=True)
|
||||
self.parent, self.child = Pipe()
|
||||
self.chatglm_model = None
|
||||
self.chatglm_tokenizer = None
|
||||
self.info = ""
|
||||
self.success = True
|
||||
self.check_dependency()
|
||||
self.start()
|
||||
self.threadLock = threading.Lock()
|
||||
|
||||
def check_dependency(self):
|
||||
try:
|
||||
import sentencepiece
|
||||
self.info = "依赖检测通过"
|
||||
self.success = True
|
||||
except:
|
||||
self.info = "缺少ChatGLM的依赖，如果要使用ChatGLM，除了基础的pip依赖以外，您还需要运行`pip install -r request_llms/requirements_chatglm.txt`安装ChatGLM的依赖。"
|
||||
self.success = False
|
||||
# ------------------------------------------------------------------------------------------------------------------------
|
||||
# 🔌💻 Local Model
|
||||
# ------------------------------------------------------------------------------------------------------------------------
|
||||
@SingletonLocalLLM
|
||||
class GetGLM2Handle(LocalLLMHandle):
|
||||
|
||||
def ready(self):
|
||||
return self.chatglm_model is not None
|
||||
def load_model_info(self):
|
||||
# 🏃♂️🏃♂️🏃♂️ 子进程执行
|
||||
self.model_name = model_name
|
||||
self.cmd_to_install = cmd_to_install
|
||||
|
||||
def run(self):
|
||||
# 子进程执行
|
||||
# 第一次运行，加载参数
|
||||
retry = 0
|
||||
def load_model_and_tokenizer(self):
|
||||
# 🏃♂️🏃♂️🏃♂️ 子进程执行
|
||||
import os, glob
|
||||
import os
|
||||
import platform
|
||||
LOCAL_MODEL_QUANT, device = get_conf('LOCAL_MODEL_QUANT', 'LOCAL_MODEL_DEVICE')
|
||||
|
||||
if LOCAL_MODEL_QUANT == "INT4": # INT4
|
||||
@ -46,122 +33,47 @@ class GetGLMHandle(Process):
|
||||
else:
|
||||
_model_name_ = "THUDM/chatglm2-6b" # FP16
|
||||
|
||||
while True:
|
||||
try:
|
||||
with ProxyNetworkActivate('Download_LLM'):
|
||||
if self.chatglm_model is None:
|
||||
self.chatglm_tokenizer = AutoTokenizer.from_pretrained(_model_name_, trust_remote_code=True)
|
||||
if device=='cpu':
|
||||
self.chatglm_model = AutoModel.from_pretrained(_model_name_, trust_remote_code=True).float()
|
||||
else:
|
||||
self.chatglm_model = AutoModel.from_pretrained(_model_name_, trust_remote_code=True).half().cuda()
|
||||
self.chatglm_model = self.chatglm_model.eval()
|
||||
break
|
||||
else:
|
||||
break
|
||||
except:
|
||||
retry += 1
|
||||
if retry > 3:
|
||||
self.child.send('[Local Message] Call ChatGLM fail 不能正常加载ChatGLM的参数。')
|
||||
raise RuntimeError("不能正常加载ChatGLM的参数！")
|
||||
|
||||
while True:
|
||||
# 进入任务等待状态
|
||||
kwargs = self.child.recv()
|
||||
# 收到消息，开始请求
|
||||
try:
|
||||
for response, history in self.chatglm_model.stream_chat(self.chatglm_tokenizer, **kwargs):
|
||||
self.child.send(response)
|
||||
# # 中途接收可能的终止指令（如果有的话）
|
||||
# if self.child.poll():
|
||||
# command = self.child.recv()
|
||||
# if command == '[Terminate]': break
|
||||
except:
|
||||
from toolbox import trimmed_format_exc
|
||||
self.child.send('[Local Message] Call ChatGLM fail.' + '\n```\n' + trimmed_format_exc() + '\n```\n')
|
||||
# 请求处理结束，开始下一个循环
|
||||
self.child.send('[Finish]')
|
||||
|
||||
def stream_chat(self, **kwargs):
|
||||
# 主进程执行
|
||||
self.threadLock.acquire()
|
||||
self.parent.send(kwargs)
|
||||
while True:
|
||||
res = self.parent.recv()
|
||||
if res != '[Finish]':
|
||||
yield res
|
||||
with ProxyNetworkActivate('Download_LLM'):
|
||||
chatglm_tokenizer = AutoTokenizer.from_pretrained(_model_name_, trust_remote_code=True)
|
||||
if device=='cpu':
|
||||
chatglm_model = AutoModel.from_pretrained(_model_name_, trust_remote_code=True).float()
|
||||
else:
|
||||
break
|
||||
self.threadLock.release()
|
||||
chatglm_model = AutoModel.from_pretrained(_model_name_, trust_remote_code=True).half().cuda()
|
||||
chatglm_model = chatglm_model.eval()
|
||||
|
||||
global glm_handle
|
||||
glm_handle = None
|
||||
#################################################################################
|
||||
def predict_no_ui_long_connection(inputs, llm_kwargs, history=[], sys_prompt="", observe_window=[], console_slience=False):
|
||||
"""
|
||||
多线程方法
|
||||
函数的说明请见 request_llms/bridge_all.py
|
||||
"""
|
||||
global glm_handle
|
||||
if glm_handle is None:
|
||||
glm_handle = GetGLMHandle()
|
||||
if len(observe_window) >= 1: observe_window[0] = load_message + "\n\n" + glm_handle.info
|
||||
if not glm_handle.success:
|
||||
error = glm_handle.info
|
||||
glm_handle = None
|
||||
raise RuntimeError(error)
|
||||
self._model = chatglm_model
|
||||
self._tokenizer = chatglm_tokenizer
|
||||
return self._model, self._tokenizer
|
||||
|
||||
# chatglm 没有 sys_prompt 接口，因此把prompt加入 history
|
||||
history_feedin = []
|
||||
history_feedin.append(["What can I do?", sys_prompt])
|
||||
for i in range(len(history)//2):
|
||||
history_feedin.append([history[2*i], history[2*i+1]] )
|
||||
def llm_stream_generator(self, **kwargs):
|
||||
# 🏃♂️🏃♂️🏃♂️ 子进程执行
|
||||
def adaptor(kwargs):
|
||||
query = kwargs['query']
|
||||
max_length = kwargs['max_length']
|
||||
top_p = kwargs['top_p']
|
||||
temperature = kwargs['temperature']
|
||||
history = kwargs['history']
|
||||
return query, max_length, top_p, temperature, history
|
||||
|
||||
watch_dog_patience = 5 # 看门狗 (watchdog) 的耐心, 设置5秒即可
|
||||
response = ""
|
||||
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']):
|
||||
if len(observe_window) >= 1: observe_window[0] = response
|
||||
if len(observe_window) >= 2:
|
||||
if (time.time()-observe_window[1]) > watch_dog_patience:
|
||||
raise RuntimeError("程序终止。")
|
||||
return response
|
||||
query, max_length, top_p, temperature, history = adaptor(kwargs)
|
||||
|
||||
for response, history in self._model.stream_chat(self._tokenizer,
|
||||
query,
|
||||
history,
|
||||
max_length=max_length,
|
||||
top_p=top_p,
|
||||
temperature=temperature,
|
||||
):
|
||||
yield response
|
||||
|
||||
def try_to_import_special_deps(self, **kwargs):
|
||||
# import something that will raise error if the user does not install requirement_*.txt
|
||||
# 🏃♂️🏃♂️🏃♂️ 主进程执行
|
||||
import importlib
|
||||
# importlib.import_module('modelscope')
|
||||
|
||||
|
||||
|
||||
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_llms/bridge_all.py
|
||||
"""
|
||||
chatbot.append((inputs, ""))
|
||||
|
||||
global glm_handle
|
||||
if glm_handle is None:
|
||||
glm_handle = GetGLMHandle()
|
||||
chatbot[-1] = (inputs, load_message + "\n\n" + glm_handle.info)
|
||||
yield from update_ui(chatbot=chatbot, history=[])
|
||||
if not glm_handle.success:
|
||||
glm_handle = None
|
||||
return
|
||||
|
||||
if additional_fn is not None:
|
||||
from core_functional import handle_core_functionality
|
||||
inputs, history = handle_core_functionality(additional_fn, inputs, history, chatbot)
|
||||
|
||||
# 处理历史信息
|
||||
history_feedin = []
|
||||
history_feedin.append(["What can I do?", system_prompt] )
|
||||
for i in range(len(history)//2):
|
||||
history_feedin.append([history[2*i], history[2*i+1]] )
|
||||
|
||||
# 开始接收chatglm的回复
|
||||
response = "[Local Message] 等待ChatGLM响应中 ..."
|
||||
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)
|
||||
|
||||
# 总结输出
|
||||
if response == "[Local Message] 等待ChatGLM响应中 ...":
|
||||
response = "[Local Message] ChatGLM响应异常 ..."
|
||||
history.extend([inputs, response])
|
||||
yield from update_ui(chatbot=chatbot, history=history)
|
||||
# ------------------------------------------------------------------------------------------------------------------------
|
||||
# 🔌💻 GPT-Academic Interface
|
||||
# ------------------------------------------------------------------------------------------------------------------------
|
||||
predict_no_ui_long_connection, predict = get_local_llm_predict_fns(GetGLM2Handle, model_name)
|
||||
@ -12,7 +12,7 @@ from .local_llm_class import LocalLLMHandle, get_local_llm_predict_fns, Singleto
|
||||
# 🔌💻 Local Model
|
||||
# ------------------------------------------------------------------------------------------------------------------------
|
||||
@SingletonLocalLLM
|
||||
class GetONNXGLMHandle(LocalLLMHandle):
|
||||
class GetGLM3Handle(LocalLLMHandle):
|
||||
|
||||
def load_model_info(self):
|
||||
# 🏃♂️🏃♂️🏃♂️ 子进程执行
|
||||
@ -69,10 +69,10 @@ class GetONNXGLMHandle(LocalLLMHandle):
|
||||
# import something that will raise error if the user does not install requirement_*.txt
|
||||
# 🏃♂️🏃♂️🏃♂️ 主进程执行
|
||||
import importlib
|
||||
importlib.import_module('modelscope')
|
||||
# importlib.import_module('modelscope')
|
||||
|
||||
|
||||
# ------------------------------------------------------------------------------------------------------------------------
|
||||
# 🔌💻 GPT-Academic Interface
|
||||
# ------------------------------------------------------------------------------------------------------------------------
|
||||
predict_no_ui_long_connection, predict = get_local_llm_predict_fns(GetONNXGLMHandle, model_name, history_format='chatglm3')
|
||||
predict_no_ui_long_connection, predict = get_local_llm_predict_fns(GetGLM3Handle, model_name, history_format='chatglm3')
|
||||
@ -1,15 +1,16 @@
|
||||
from transformers import AutoModel, AutoTokenizer
|
||||
import time
|
||||
import threading
|
||||
import importlib
|
||||
from toolbox import update_ui, get_conf, Singleton
|
||||
from toolbox import update_ui
|
||||
from multiprocessing import Process, Pipe
|
||||
from contextlib import redirect_stdout
|
||||
|
||||
|
||||
def SingletonLocalLLM(cls):
|
||||
"""
|
||||
一个单实例装饰器
|
||||
"""
|
||||
_instance = {}
|
||||
|
||||
def _singleton(*args, **kargs):
|
||||
if cls not in _instance:
|
||||
_instance[cls] = cls(*args, **kargs)
|
||||
@ -21,6 +22,28 @@ def SingletonLocalLLM(cls):
|
||||
return _instance[cls]
|
||||
return _singleton
|
||||
|
||||
|
||||
def reset_tqdm_output():
|
||||
import sys, tqdm
|
||||
def status_printer(self, file):
|
||||
fp = file
|
||||
if fp in (sys.stderr, sys.stdout):
|
||||
getattr(sys.stderr, 'flush', lambda: None)()
|
||||
getattr(sys.stdout, 'flush', lambda: None)()
|
||||
|
||||
def fp_write(s):
|
||||
print(s)
|
||||
last_len = [0]
|
||||
|
||||
def print_status(s):
|
||||
from tqdm.utils import disp_len
|
||||
len_s = disp_len(s)
|
||||
fp_write('\r' + s + (' ' * max(last_len[0] - len_s, 0)))
|
||||
last_len[0] = len_s
|
||||
return print_status
|
||||
tqdm.tqdm.status_printer = status_printer
|
||||
|
||||
|
||||
class LocalLLMHandle(Process):
|
||||
def __init__(self):
|
||||
# ⭐主进程执行
|
||||
@ -28,6 +51,9 @@ class LocalLLMHandle(Process):
|
||||
self.corrupted = False
|
||||
self.load_model_info()
|
||||
self.parent, self.child = Pipe()
|
||||
# allow redirect_stdout
|
||||
self.std_tag = "[Subprocess Message] "
|
||||
self.child.write = lambda x: self.child.send(self.std_tag + x)
|
||||
self.running = True
|
||||
self._model = None
|
||||
self._tokenizer = None
|
||||
@ -64,7 +90,7 @@ class LocalLLMHandle(Process):
|
||||
# ⭐主进程执行
|
||||
try:
|
||||
self.try_to_import_special_deps()
|
||||
self.info = "依赖检测通过"
|
||||
self.info = "`依赖检测通过`"
|
||||
self.running = True
|
||||
except:
|
||||
self.info = f"缺少{self.model_name}的依赖，如果要使用{self.model_name}，除了基础的pip依赖以外，您还需要运行{self.cmd_to_install}安装{self.model_name}的依赖。"
|
||||
@ -73,15 +99,21 @@ class LocalLLMHandle(Process):
|
||||
def run(self):
|
||||
# 🏃♂️🏃♂️🏃♂️ 子进程执行
|
||||
# 第一次运行，加载参数
|
||||
reset_tqdm_output()
|
||||
self.info = "`尝试加载模型`"
|
||||
try:
|
||||
self._model, self._tokenizer = self.load_model_and_tokenizer()
|
||||
with redirect_stdout(self.child):
|
||||
self._model, self._tokenizer = self.load_model_and_tokenizer()
|
||||
except:
|
||||
self.info = "`加载模型失败`"
|
||||
self.running = False
|
||||
from toolbox import trimmed_format_exc
|
||||
self.child.send(f'[Local Message] 不能正常加载{self.model_name}的参数.' + '\n```\n' + trimmed_format_exc() + '\n```\n')
|
||||
self.child.send(
|
||||
f'[Local Message] 不能正常加载{self.model_name}的参数.' + '\n```\n' + trimmed_format_exc() + '\n```\n')
|
||||
self.child.send('[FinishBad]')
|
||||
raise RuntimeError(f"不能正常加载{self.model_name}的参数！")
|
||||
|
||||
self.info = "`准备就绪`"
|
||||
while True:
|
||||
# 进入任务等待状态
|
||||
kwargs = self.child.recv()
|
||||
@ -93,25 +125,35 @@ class LocalLLMHandle(Process):
|
||||
# 请求处理结束，开始下一个循环
|
||||
except:
|
||||
from toolbox import trimmed_format_exc
|
||||
self.child.send(f'[Local Message] 调用{self.model_name}失败.' + '\n```\n' + trimmed_format_exc() + '\n```\n')
|
||||
self.child.send(
|
||||
f'[Local Message] 调用{self.model_name}失败.' + '\n```\n' + trimmed_format_exc() + '\n```\n')
|
||||
self.child.send('[Finish]')
|
||||
|
||||
def stream_chat(self, **kwargs):
|
||||
# ⭐主进程执行
|
||||
self.threadLock.acquire()
|
||||
self.parent.send(kwargs)
|
||||
while True:
|
||||
res = self.parent.recv()
|
||||
if res == '[Finish]':
|
||||
break
|
||||
if res == '[FinishBad]':
|
||||
self.running = False
|
||||
self.corrupted = True
|
||||
break
|
||||
else:
|
||||
yield res
|
||||
self.threadLock.release()
|
||||
|
||||
if self.info == "`准备就绪`":
|
||||
yield "`正在等待线程锁，排队中请稍后 ...`"
|
||||
with self.threadLock:
|
||||
self.parent.send(kwargs)
|
||||
std_out = ""
|
||||
std_out_clip_len = 4096
|
||||
while True:
|
||||
res = self.parent.recv()
|
||||
if res.startswith(self.std_tag):
|
||||
new_output = res[len(self.std_tag):]
|
||||
std_out = std_out[:std_out_clip_len]
|
||||
print(new_output, end='')
|
||||
std_out = new_output + std_out
|
||||
yield self.std_tag + '\n```\n' + std_out + '\n```\n'
|
||||
elif res == '[Finish]':
|
||||
break
|
||||
elif res == '[FinishBad]':
|
||||
self.running = False
|
||||
self.corrupted = True
|
||||
break
|
||||
else:
|
||||
std_out = ""
|
||||
yield res
|
||||
|
||||
|
||||
def get_local_llm_predict_fns(LLMSingletonClass, model_name, history_format='classic'):
|
||||
@ -123,15 +165,17 @@ def get_local_llm_predict_fns(LLMSingletonClass, model_name, history_format='cla
|
||||
函数的说明请见 request_llms/bridge_all.py
|
||||
"""
|
||||
_llm_handle = LLMSingletonClass()
|
||||
if len(observe_window) >= 1: observe_window[0] = load_message + "\n\n" + _llm_handle.info
|
||||
if not _llm_handle.running: raise RuntimeError(_llm_handle.info)
|
||||
if len(observe_window) >= 1:
|
||||
observe_window[0] = load_message + "\n\n" + _llm_handle.info
|
||||
if not _llm_handle.running:
|
||||
raise RuntimeError(_llm_handle.info)
|
||||
|
||||
if history_format == 'classic':
|
||||
# 没有 sys_prompt 接口，因此把prompt加入 history
|
||||
history_feedin = []
|
||||
history_feedin.append([sys_prompt, "Certainly!"])
|
||||
for i in range(len(history)//2):
|
||||
history_feedin.append([history[2*i], history[2*i+1]] )
|
||||
history_feedin.append([history[2*i], history[2*i+1]])
|
||||
elif history_format == 'chatglm3':
|
||||
# 有 sys_prompt 接口
|
||||
conversation_cnt = len(history) // 2
|
||||
@ -145,24 +189,24 @@ def get_local_llm_predict_fns(LLMSingletonClass, model_name, history_format='cla
|
||||
what_gpt_answer["role"] = "assistant"
|
||||
what_gpt_answer["content"] = history[index+1]
|
||||
if what_i_have_asked["content"] != "":
|
||||
if what_gpt_answer["content"] == "": continue
|
||||
if what_gpt_answer["content"] == "":
|
||||
continue
|
||||
history_feedin.append(what_i_have_asked)
|
||||
history_feedin.append(what_gpt_answer)
|
||||
else:
|
||||
history_feedin[-1]['content'] = what_gpt_answer['content']
|
||||
|
||||
watch_dog_patience = 5 # 看门狗 (watchdog) 的耐心, 设置5秒即可
|
||||
watch_dog_patience = 5 # 看门狗 (watchdog) 的耐心, 设置5秒即可
|
||||
response = ""
|
||||
for response in _llm_handle.stream_chat(query=inputs, history=history_feedin, max_length=llm_kwargs['max_length'], top_p=llm_kwargs['top_p'], temperature=llm_kwargs['temperature']):
|
||||
if len(observe_window) >= 1:
|
||||
observe_window[0] = response
|
||||
if len(observe_window) >= 2:
|
||||
if (time.time()-observe_window[1]) > watch_dog_patience: raise RuntimeError("程序终止。")
|
||||
if (time.time()-observe_window[1]) > watch_dog_patience:
|
||||
raise RuntimeError("程序终止。")
|
||||
return response
|
||||
|
||||
|
||||
|
||||
def predict(inputs, llm_kwargs, plugin_kwargs, chatbot, history=[], system_prompt='', stream = True, additional_fn=None):
|
||||
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_llms/bridge_all.py
|
||||
@ -172,11 +216,13 @@ def get_local_llm_predict_fns(LLMSingletonClass, model_name, history_format='cla
|
||||
_llm_handle = LLMSingletonClass()
|
||||
chatbot[-1] = (inputs, load_message + "\n\n" + _llm_handle.info)
|
||||
yield from update_ui(chatbot=chatbot, history=[])
|
||||
if not _llm_handle.running: raise RuntimeError(_llm_handle.info)
|
||||
if not _llm_handle.running:
|
||||
raise RuntimeError(_llm_handle.info)
|
||||
|
||||
if additional_fn is not None:
|
||||
from core_functional import handle_core_functionality
|
||||
inputs, history = handle_core_functionality(additional_fn, inputs, history, chatbot)
|
||||
inputs, history = handle_core_functionality(
|
||||
additional_fn, inputs, history, chatbot)
|
||||
|
||||
# 处理历史信息
|
||||
if history_format == 'classic':
|
||||
@ -184,7 +230,7 @@ def get_local_llm_predict_fns(LLMSingletonClass, model_name, history_format='cla
|
||||
history_feedin = []
|
||||
history_feedin.append([system_prompt, "Certainly!"])
|
||||
for i in range(len(history)//2):
|
||||
history_feedin.append([history[2*i], history[2*i+1]] )
|
||||
history_feedin.append([history[2*i], history[2*i+1]])
|
||||
elif history_format == 'chatglm3':
|
||||
# 有 sys_prompt 接口
|
||||
conversation_cnt = len(history) // 2
|
||||
@ -198,7 +244,8 @@ def get_local_llm_predict_fns(LLMSingletonClass, model_name, history_format='cla
|
||||
what_gpt_answer["role"] = "assistant"
|
||||
what_gpt_answer["content"] = history[index+1]
|
||||
if what_i_have_asked["content"] != "":
|
||||
if what_gpt_answer["content"] == "": continue
|
||||
if what_gpt_answer["content"] == "":
|
||||
continue
|
||||
history_feedin.append(what_i_have_asked)
|
||||
history_feedin.append(what_gpt_answer)
|
||||
else:
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user