记录微调ChatGLM3_attributeerror: can't set attribute 'eos_token

原文章链接

记录微调chatglm3

遇到的问题：

1.在合并模型阶段NotImplementedError: Cannot copy out of meta tensor; no data!

解决方案：
新建一个SCRIPT.py文件内容如下：

# Copyright (c) OpenMMLab. All rights reserved.
import argparse

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer


def parse_args():
    parser = argparse.ArgumentParser(
        description='Merge a HuggingFace adapter to LLM')
    parser.add_argument('model_name_or_path', help='model name or path')
    parser.add_argument('adapter_name_or_path', help='adapter name or path')
    parser.add_argument(
        'save_dir', help='the directory to save the merged model')
    parser.add_argument(
        '--max-shard-size',
        type=str,
        default='2GB',
        help='Only applicable for LLM. The maximum size for '
        'each sharded checkpoint.')
    parser.add_argument(
        '--offload-folder',
        default=None,
        help='The folder in which to offload the model weights (or where '
        'the model weights are already offloaded).')
    args = parser.parse_args()
    return args


def main():
    args = parse_args()
    model = AutoModelForCausalLM.from_pretrained(
        args.model_name_or_path,
        torch_dtype=torch.float16,
        low_cpu_mem_usage=False,   #如果报错NotImplementedError: Cannot copy out of meta tensor; no data! ，将这个函数设置为False并且注释掉device_map
        # device_map='auto',
        offload_folder=args.offload_folder,
        trust_remote_code=True,
        empty_init=False)
    tokenizer = AutoTokenizer.from_pretrained(
        args.model_name_or_path,
        trust_remote_code=True,
        encode_special_tokens=True)
    model_unmerged = PeftModel.from_pretrained(
        model,
        args.adapter_name_or_path,
        device_map='auto',
        torch_dtype=torch.float16,
        offload_folder=args.offload_folder,
        is_trainable=False)
    model_merged = model_unmerged.merge_and_unload()
    print(f'Saving to {args.save_dir}...')
    model_merged.save_pretrained(
        args.save_dir, max_shard_size=args.max_shard_size)
    tokenizer.save_pretrained(args.save_dir)
    print('All done!')


if __name__ == '__main__':
    main()

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

然后运行

python SCRIPT.py 原始模型路径 hf模型保存路径 合并后你想保存的路径
1

2.在预测阶段AttributeError: can’t set attribute ‘eos_token’

把源目录除了 bin 和 pytorch_model.bin.index.json 以外的文件全部复制到导出目录中覆盖
1