Ziya-LLaMA-13B 模型在GPU 上部署

Ziya-LLaMA-13B 模型在GPU 上部署

Ziya-LLaMA-13B是IDEA-CCNL基于LLaMa的130亿参数的大规模预训练模型,具备翻译,编程,文本分类,信息抽取,摘要,文案生成,常识问答和数学计算等能力。目前姜子牙通用大模型已完成大规模预训练、多任务有监督微调和人类反馈学习三阶段的训练过程。

1. 部署准备

1.1 硬件环境
显卡最低显存为54GB,可以为一张A100(80GB)或者两张A100(40GB)。
6.10日更新增加了模型量化模型,现在的模型只需要一张A100或者两张A10(40GB显卡即可推理)

1.2 python环境

  cpm_kernels==1.0.11
  gradio==3.34.0
  huggingface-hub==0.16.4
  Pillow==9.5.0
  torch-1.12.1+cu113-cp38-cp38-linux_x86_64.whl (https://download.pytorch.org/whl/torch/)
  tqdm==4.64.1
  transformers==4.31.0
  accelerate==0.20.3
   

1.3 模型下载

2 . 模型部署

首先说明一下项目的文件系统目录

  -llama-13B
  -llama-13B-convert
  -ziya_v1.1_delta
  -ziya_v1.1
  -apply_delta.py
  -convert_llama_weights_to_hf.py
  -launch.py
  -utils.py

其中

  • llama-13B为llama原始参数存放的目录,
  • llama-13B-convert为转换成huggingface形式的参数存放的目录,
  • ziya_v1.1_delta为huggingface上的权重文件,
  • ziya_v1.1为最终转换后的权重文件。
  • launch.py为本地化部署文件,

详见后续章节,utils.py为官方给的文件,直接从https://modelscope.cn/studios/Fengshenbang/Ziya_LLaMA_13B_v1_online/files下载即可。

llama-13B为llama原始参数存放的目录, 原始模型权重不太好下载,可以不用管

llama-13B-convert为转换成huggingface形式的参数存放的目录, 可以直接从网上找转化好的模型权重数据

https://huggingface.co/decapoda-research/llama-13b-hf/tree/main
https://pan.baidu.com/s/1Y7YWdFWX1Yzy2Yuujp8Tqg?pwd=1p5n#list/path=%2FLLaMA(1)%2Fllama-13b

ziya_v1.1_delta为huggingface上的权重文件
https://huggingface.co/IDEA-CCNL/Ziya-LLaMA-13B-v1.1/tree/main

2.1 llama-13B权重转换(如直接下载转化好的模型权重数据,该步骤可省略)

首先第一步需要将llama-13B的原始权重转换成huggingface的权重形式,使用convert_llama_weights_to_hf.py脚本进行转换,转换代码如下:

  python convert_llama_weights_to_hf.py --input_dir $你的llama-13B路径 --model_size 13B --output_dir $你的llama-13B模型转换后的路径

2.2 结合基础的llama权重和Ziya-LLaMA-13B delta权重得到Ziya-LLaMA-13B权重

使用如下代码得到最终的Ziya-LLaMA-13B权重。

  python -m apply_delta --base $你的llama-13B模型转换后的路径 --target $你的最终权重存储路径 --delta $你下载的Ziya-LLaMA-13B权重路径

2.3 模型部署

使用下面这段代码进行模型部署。该代码修改自https://modelscope.cn/studios/Fengshenbang/Ziya_LLaMA_13B_v1_online/files在官方代码的基础上增加了一些参数和优化内容。

  import gradio as gr
  import os
  import gc
  import torch
   
   
  from transformers import AutoTokenizer
  os.environ['CUDA_VISIBLE_DEVICES'] = "0,1"
  from utils import SteamGenerationMixin
   
   
  class MindBot(object):
  def __init__(self):
  model_path = './ziya_v1.1'
  self.model = SteamGenerationMixin.from_pretrained(model_path, device_map='auto').half()
  self.model.eval()
   
  self.tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False)
   
  def build_prompt(self, instruction, history, human='<human>', bot='<bot>'):
  pmt = ''
  if len(history) > 0:
  for line in history:
  pmt += f'{human}: {line[0].strip()}\n{bot}: {line[1]}\n'
  pmt += f'{human}: {instruction.strip()}\n{bot}: \n'
  return pmt
   
  def interaction(
  self,
  instruction,
  history,
  max_new_tokens,
  temperature,
  top_p,
  max_memory=1024
  ):
   
  prompt = self.build_prompt(instruction, history)
  input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids
  if input_ids.shape[1] > max_memory:
  input_ids = input_ids[:, -max_memory:]
   
  prompt_len = input_ids.shape[1]
  # stream generation method
  try:
  tmp = history.copy()
  output = ''
  with torch.no_grad():
  for generation_output in self.model.stream_generate(
  input_ids.cuda(),
  max_new_tokens=max_new_tokens,
  do_sample=True,
  top_p=top_p,
  temperature=temperature,
  repetition_penalty=1.,
  eos_token_id=2,
  bos_token_id=1,
  pad_token_id=0
  ):
  s = generation_output[0][prompt_len:]
  output = self.tokenizer.decode(s, skip_special_tokens=True)
  # output = output.replace('\n', '<br>')
  output = output.replace('\n', '\n\n')
  tmp.append((instruction, output))
  yield '', tmp
  tmp.pop()
  # gc.collect()
  # torch.cuda.empty_cache()
  history.append((instruction, output))
  print('input -----> \n', prompt)
  print('output -------> \n', output)
  print('history: ======> \n', history)
  except torch.cuda.OutOfMemoryError:
  gc.collect()
  torch.cuda.empty_cache()
  self.model.empty_cache()
  history.append((instruction, "【显存不足,请清理历史信息后再重试】"))
  return "", history
   
  def chat(self):
   
  with gr.Blocks(title='IDEA MindBot', css=".bgcolor {color: white !important; background: #FFA500 !important;}") as demo:
  with gr.Row():
  gr.Column(scale=0.25)
  with gr.Column(scale=0.5):
  gr.Markdown("<center><h1>IDEA Ziya</h1></center>")
  gr.Markdown("<center>姜子牙通用大模型V1.1是基于LLaMa的130亿参数的大规模预训练模型,具备翻译,编程,文本分类,信息抽取,摘要,文案生成,常识问答和数学计算等能力。目前姜子牙通用大模型已完成大规模预训练、多任务有监督微调和人类反馈学习三阶段的训练过程。</center>")
  gr.Column(scale=0.25)
  with gr.Row():
  gr.Column(scale=0.25)
  with gr.Column(scale=0.5):
  chatbot = gr.Chatbot(label='Ziya').style(height=500)
  msg = gr.Textbox(label="Input")
  # gr.Column(scale=0.25)
  with gr.Column(scale=0.25):
  max_new_tokens = gr.Slider(0, 2048, value=1024, step=1.0, label="Max_new_tokens", interactive=True)
  top_p = gr.Slider(0, 1, value=0.85, step=0.01, label="Top P", interactive=True)
  temperature = gr.Slider(0, 1, value=0.8, step=0.01, label="Temperature", interactive=True)
  with gr.Row():
  gr.Column(scale=0.25)
  with gr.Column(scale=0.25):
  clear = gr.Button("Clear")
  with gr.Column(scale=0.25):
  submit = gr.Button("Submit")
  gr.Column(scale=0.25)
   
  msg.submit(self.interaction, [msg, chatbot,max_new_tokens,top_p,temperature], [msg, chatbot])
  clear.click(lambda: None, None, chatbot, queue=False)
  submit.click(self.interaction, [msg, chatbot,max_new_tokens,top_p,temperature], [msg, chatbot])
  return demo.queue(concurrency_count=10).launch(share=True,server_name="127.0.0.1", server_port=7886)
   
   
  if __name__ == '__main__':
  mind_bot = MindBot()
  mind_bot.chat()
   
   

2张A10 响应还是有点慢
2张A10 显存都快爆掉了

FAQ:

  RuntimeError: Failed to import transformers.models.clipseg.modeling_clipseg because of the following error (look up to see its traceback):
  /usr/local/lib/python3.8/dist-packages/transformer_engine_extensions.cpython-38-x86_64-linux-gnu.so: undefined symbol: _ZN2at4_ops5zeros4callEN3c108ArrayRefINS2_6SymIntEEENS2_8optionalINS2_10ScalarTypeEEENS6_INS2_6LayoutEEENS6_INS2_6DeviceEEENS6_IbEE
   
VBNET 复制 全屏

pip uninstall transformer-engine 可解决

参考文档:

https://github.com/ChaosWang666/Ziya-LLaMA-13B-deployment
https://welearnnlp.blog.csdn.net/article/details/131125481

https://www.cnblogs.com/Flat-White/p/17048075.html
https://download.pytorch.org/whl/cu113
https://download.pytorch.org/whl/torch/
https://download.pytorch.org/whl/torchvision/

https://github.com/ChaosWang666/Ziya-LLaMA-13B-deployment/blob/main/apply_delta.py#L148

 
标签: GPU , Ziya-LLaMA , 大模型
 

posted on 2024-11-19 19:33  漫思  阅读(9)  评论(0编辑  收藏  举报

导航