gpt-academic / crazy_functions /chatglm微调工具.py
qingxu98's picture
update
d0703ef
raw
history blame
6.39 kB
from toolbox import CatchException, update_ui, promote_file_to_downloadzone
from .crazy_utils import request_gpt_model_multi_threads_with_very_awesome_ui_and_high_efficiency
import datetime, json
def fetch_items(list_of_items, batch_size):
for i in range(0, len(list_of_items), batch_size):
yield list_of_items[i:i + batch_size]
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)
parser.add_argument("--pre_seq_len", type=int, help="pre_seq_len", default=50)
parser.add_argument("--learning_rate", type=float, help="learning_rate", default=2e-2)
parser.add_argument("--num_gpus", type=int, help="num_gpus", default=1)
parser.add_argument("--json_dataset", type=str, help="json_dataset", default="")
parser.add_argument("--ptuning_directory", type=str, help="ptuning_directory", default="")
# Parse the arguments
args = parser.parse_args(shlex.split(arguments))
return args
@CatchException
def 微调数据集生成(txt, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt, user_request):
"""
txt 输入栏用户输入的文本,例如需要翻译的一段话,再例如一个包含了待处理文件的路径
llm_kwargs gpt模型参数,如温度和top_p等,一般原样传递下去就行
plugin_kwargs 插件模型的参数
chatbot 聊天显示框的句柄,用于显示给用户
history 聊天历史,前情提要
system_prompt 给gpt的静默提醒
user_request 当前用户的请求信息(IP地址等)
"""
history = [] # 清空历史,以免输入溢出
chatbot.append(("这是什么功能?", "[Local Message] 微调数据集生成"))
if ("advanced_arg" in plugin_kwargs) and (plugin_kwargs["advanced_arg"] == ""): plugin_kwargs.pop("advanced_arg")
args = plugin_kwargs.get("advanced_arg", None)
if args is None:
chatbot.append(("没给定指令", "退出"))
yield from update_ui(chatbot=chatbot, history=history); return
else:
arguments = string_to_options(arguments=args)
dat = []
with open(txt, 'r', encoding='utf8') as f:
for line in f.readlines():
json_dat = json.loads(line)
dat.append(json_dat["content"])
llm_kwargs['llm_model'] = arguments.llm_to_learn
for batch in fetch_items(dat, arguments.batch):
res = yield from request_gpt_model_multi_threads_with_very_awesome_ui_and_high_efficiency(
inputs_array=[f"{arguments.prompt_prefix}\n\n{b}" for b in (batch)],
inputs_show_user_array=[f"Show Nothing" for _ in (batch)],
llm_kwargs=llm_kwargs,
chatbot=chatbot,
history_array=[[] for _ in (batch)],
sys_prompt_array=[arguments.system_prompt for _ in (batch)],
max_workers=10 # OpenAI所允许的最大并行过载
)
with open(txt+'.generated.json', 'a+', encoding='utf8') as f:
for b, r in zip(batch, res[1::2]):
f.write(json.dumps({"content":b, "summary":r}, ensure_ascii=False)+'\n')
promote_file_to_downloadzone(txt+'.generated.json', rename_file='generated.json', chatbot=chatbot)
return
@CatchException
def 启动微调(txt, llm_kwargs, plugin_kwargs, chatbot, history, system_prompt, user_request):
"""
txt 输入栏用户输入的文本,例如需要翻译的一段话,再例如一个包含了待处理文件的路径
llm_kwargs gpt模型参数,如温度和top_p等,一般原样传递下去就行
plugin_kwargs 插件模型的参数
chatbot 聊天显示框的句柄,用于显示给用户
history 聊天历史,前情提要
system_prompt 给gpt的静默提醒
user_request 当前用户的请求信息(IP地址等)
"""
import subprocess
history = [] # 清空历史,以免输入溢出
chatbot.append(("这是什么功能?", "[Local Message] 微调数据集生成"))
if ("advanced_arg" in plugin_kwargs) and (plugin_kwargs["advanced_arg"] == ""): plugin_kwargs.pop("advanced_arg")
args = plugin_kwargs.get("advanced_arg", None)
if args is None:
chatbot.append(("没给定指令", "退出"))
yield from update_ui(chatbot=chatbot, history=history); return
else:
arguments = string_to_options(arguments=args)
pre_seq_len = arguments.pre_seq_len # 128
learning_rate = arguments.learning_rate # 2e-2
num_gpus = arguments.num_gpus # 1
json_dataset = arguments.json_dataset # 't_code.json'
ptuning_directory = arguments.ptuning_directory # '/home/hmp/ChatGLM2-6B/ptuning'
command = f"torchrun --standalone --nnodes=1 --nproc-per-node={num_gpus} main.py \
--do_train \
--train_file AdvertiseGen/{json_dataset} \
--validation_file AdvertiseGen/{json_dataset} \
--preprocessing_num_workers 20 \
--prompt_column content \
--response_column summary \
--overwrite_cache \
--model_name_or_path THUDM/chatglm2-6b \
--output_dir output/clothgen-chatglm2-6b-pt-{pre_seq_len}-{learning_rate} \
--overwrite_output_dir \
--max_source_length 256 \
--max_target_length 256 \
--per_device_train_batch_size 1 \
--per_device_eval_batch_size 1 \
--gradient_accumulation_steps 16 \
--predict_with_generate \
--max_steps 100 \
--logging_steps 10 \
--save_steps 20 \
--learning_rate {learning_rate} \
--pre_seq_len {pre_seq_len} \
--quantization_bit 4"
process = subprocess.Popen(command, shell=True, cwd=ptuning_directory)
try:
process.communicate(timeout=3600*24)
except subprocess.TimeoutExpired:
process.kill()
return