Spaces:
Sleeping
Sleeping
File size: 9,415 Bytes
506c93a |
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 |
from transformers import AutoModel, AutoTokenizer
import time
import os
import json
import threading
import importlib
from toolbox import update_ui, get_conf
from multiprocessing import Process, Pipe
load_message = "ChatGLMFT尚未加载，加载需要一段时间。注意，取决于`config.py`的配置，ChatGLMFT消耗大量的内存（CPU）或显存（GPU），也许会导致低配计算机卡死 ……"
def string_to_options(arguments):
import argparse
import shlex
# Create an argparse.ArgumentParser instance
parser = argparse.ArgumentParser()
# Add command-line arguments
parser.add_argument("--llm_to_learn", type=str, help="LLM model to learn", default="gpt-3.5-turbo")
parser.add_argument("--prompt_prefix", type=str, help="Prompt prefix", default='')
parser.add_argument("--system_prompt", type=str, help="System prompt", default='')
parser.add_argument("--batch", type=int, help="System prompt", default=50)
# Parse the arguments
args = parser.parse_args(shlex.split(arguments))
return args
#################################################################################
class GetGLMFTHandle(Process):
def __init__(self):
super().__init__(daemon=True)
self.parent, self.child = Pipe()
self.chatglmft_model = None
self.chatglmft_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ChatGLMFT的依赖，如果要使用ChatGLMFT，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
def ready(self):
return self.chatglmft_model is not None
def run(self):
# 子进程执行
# 第一次运行，加载参数
retry = 0
while True:
try:
if self.chatglmft_model is None:
from transformers import AutoConfig
import torch
# conf = 'request_llms/current_ptune_model.json'
# if not os.path.exists(conf): raise RuntimeError('找不到微调模型信息')
# with open(conf, 'r', encoding='utf8') as f:
# model_args = json.loads(f.read())
CHATGLM_PTUNING_CHECKPOINT = get_conf('CHATGLM_PTUNING_CHECKPOINT')
assert os.path.exists(CHATGLM_PTUNING_CHECKPOINT), "找不到微调模型检查点"
conf = os.path.join(CHATGLM_PTUNING_CHECKPOINT, "config.json")
with open(conf, 'r', encoding='utf8') as f:
model_args = json.loads(f.read())
if 'model_name_or_path' not in model_args:
model_args['model_name_or_path'] = model_args['_name_or_path']
self.chatglmft_tokenizer = AutoTokenizer.from_pretrained(
model_args['model_name_or_path'], trust_remote_code=True)
config = AutoConfig.from_pretrained(
model_args['model_name_or_path'], trust_remote_code=True)
config.pre_seq_len = model_args['pre_seq_len']
config.prefix_projection = model_args['prefix_projection']
print(f"Loading prefix_encoder weight from {CHATGLM_PTUNING_CHECKPOINT}")
model = AutoModel.from_pretrained(model_args['model_name_or_path'], config=config, trust_remote_code=True)
prefix_state_dict = torch.load(os.path.join(CHATGLM_PTUNING_CHECKPOINT, "pytorch_model.bin"))
new_prefix_state_dict = {}
for k, v in prefix_state_dict.items():
if k.startswith("transformer.prefix_encoder."):
new_prefix_state_dict[k[len("transformer.prefix_encoder."):]] = v
model.transformer.prefix_encoder.load_state_dict(new_prefix_state_dict)
if model_args['quantization_bit'] is not None and model_args['quantization_bit'] != 0:
print(f"Quantized to {model_args['quantization_bit']} bit")
model = model.quantize(model_args['quantization_bit'])
model = model.cuda()
if model_args['pre_seq_len'] is not None:
# P-tuning v2
model.transformer.prefix_encoder.float()
self.chatglmft_model = model.eval()
break
else:
break
except Exception as e:
retry += 1
if retry > 3:
self.child.send('[Local Message] Call ChatGLMFT fail 不能正常加载ChatGLMFT的参数。')
raise RuntimeError("不能正常加载ChatGLMFT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.chatglmft_model.stream_chat(self.chatglmft_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 ChatGLMFT 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
else:
break
self.threadLock.release()
global glmft_handle
glmft_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 glmft_handle
if glmft_handle is None:
glmft_handle = GetGLMFTHandle()
if len(observe_window) >= 1: observe_window[0] = load_message + "\n\n" + glmft_handle.info
if not glmft_handle.success:
error = glmft_handle.info
glmft_handle = None
raise RuntimeError(error)
# chatglmft 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]] )
watch_dog_patience = 5 # 看门狗 (watchdog) 的耐心, 设置5秒即可
response = ""
for response in glmft_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
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 glmft_handle
if glmft_handle is None:
glmft_handle = GetGLMFTHandle()
chatbot[-1] = (inputs, load_message + "\n\n" + glmft_handle.info)
yield from update_ui(chatbot=chatbot, history=[])
if not glmft_handle.success:
glmft_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chatglmft的回复
response = "[Local Message] 等待ChatGLMFT响应中 ..."
for response in glmft_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ChatGLMFT响应中 ...":
response = "[Local Message] ChatGLMFT响应异常 ..."
history.extend([inputs, response])
yield from update_ui(chatbot=chatbot, history=history)
|