-
Notifications
You must be signed in to change notification settings - Fork 394
/
train_llms.py
191 lines (157 loc) · 8.78 KB
/
train_llms.py
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
# !/usr/bin/env python3
"""
==== No Bugs in code, just some Random Unexpected FEATURES ====
┌─────────────────────────────────────────────────────────────┐
│┌───┬───┬───┬───┬───┬───┬───┬───┬───┬───┬───┬───┬───┬───┬───┐│
││Esc│!1 │@2 │#3 │$4 │%5 │^6 │&7 │*8 │(9 │)0 │_- │+= │|\ │`~ ││
│├───┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴───┤│
││ Tab │ Q │ W │ E │ R │ T │ Y │ U │ I │ O │ P │{[ │}] │ BS ││
│├─────┴┬──┴┬──┴┬──┴┬──┴┬──┴┬──┴┬──┴┬──┴┬──┴┬──┴┬──┴┬──┴─────┤│
││ Ctrl │ A │ S │ D │ F │ G │ H │ J │ K │ L │: ;│" '│ Enter ││
│├──────┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴─┬─┴────┬───┤│
││ Shift │ Z │ X │ C │ V │ B │ N │ M │< ,│> .│? /│Shift │Fn ││
│└─────┬──┴┬──┴──┬┴───┴───┴───┴───┴───┴──┬┴───┴┬──┴┬─────┴───┘│
│ │Fn │ Alt │ Space │ Alt │Win│ HHKB │
│ └───┴─────┴───────────────────────┴─────┴───┘ │
└─────────────────────────────────────────────────────────────┘
参照 OpenLlma 改写训练文件，支持训练OpenLLAMA以外的其他模型，包含MPT等。
Authors: pankeyu
Date: 2023/05/26
Reference Authors: s-JoL(sl12160010@gmail.com)
Reference repo: Open-Llama
"""
import os
os.environ["TOKENIZERS_PARALLELISM"] = "false"
import yaml
import argparse
from rich import print
from accelerate import Accelerator
from torch.utils.data import DataLoader
from peft import LoraConfig, TaskType, get_peft_model
from datasets.distributed import split_dataset_by_node
from transformers import AutoConfig, AutoModelForCausalLM, LlamaTokenizer, AutoTokenizer
from datetime import timedelta
from accelerate.utils import InitProcessGroupKwargs
from dataset.dataset import construct_dataset
from solver.trainer import Trainer
parser = argparse.ArgumentParser()
parser.add_argument("--train_config", default=None, type=str, help="Train config path.")
parser.add_argument("--model_config", default=None, type=str, help="Model config, when you specify ckpt, this param won't be used.")
args = parser.parse_args()
with open(args.train_config, "r", encoding="utf-8") as fp:
config = yaml.load(fp, Loader=yaml.FullLoader)
print('=' * 20 + ' Current Config ' + '=' * 20)
print(config)
def main():
assert args.model_config or config["train"]["ckpt"], \
'You must specify model_config by passing --model_config or a pretrained model path by setting `ckpt` in --train_config.'
kwargs = InitProcessGroupKwargs(timeout=timedelta(seconds=3600)) # 将timeout阈值从 30min 设置为 1h
accelerator = Accelerator(
gradient_accumulation_steps=config["train"].get(
"gradient_accumulation_steps", 1
),
kwargs_handlers=[kwargs]
)
if (
'llama' in config["data"]["tokenizer_path"]
or
'Llama' in config["data"]["tokenizer_path"]
):
tokenizer = LlamaTokenizer.from_pretrained(config["data"]["tokenizer_path"])
else:
tokenizer = AutoTokenizer.from_pretrained(config["data"]["tokenizer_path"])
assert tokenizer.pad_token, \
'Detect pad_token is None, you should specify "pad_token" in tokenizer config.'
assert tokenizer.eos_token, \
'Detect eos_token is None, you should specify "eos_token" in tokenizer config.'
test_tokens = tokenizer('this is a test sentence')['input_ids'] # 测试tokenizer是否会自动添加 eos token，
config['data']['added_eos_token'] = test_tokens[-1] == tokenizer.eos_token_id # 用于后续判断是否自动为句子添加 eos token
data_config = config["data"]
if data_config.get("split_by_shard", False):
train_dataset = construct_dataset(
data_config,
tokenizer,
world_size=accelerator.num_processes
)
else:
train_dataset = construct_dataset(data_config, tokenizer)
train_dataset = split_dataset_by_node(
train_dataset,
rank=accelerator.process_index,
world_size=accelerator.num_processes,
)
train_loader = DataLoader(
train_dataset,
batch_size=config["train"]["train_batch_size"],
num_workers=config["train"]["train_num_workers"],
prefetch_factor=config["train"].get("prefetch_factor", 2),
pin_memory=True,
)
model_config = AutoConfig.from_pretrained(
args.model_config,
trust_remote_code=True
)
model_config.pad_token_id = tokenizer.pad_token_id
print('tokenizer vocab_size: ', len(tokenizer))
if config["train"]["ckpt"] is not None: # 在已有的预训练模型上进行continue training
print("Loading ckpt from: {}".format(config["train"]["ckpt"]))
raw_model = AutoModelForCausalLM.from_pretrained(
config["train"]["ckpt"],
trust_remote_code=True
)
if model_config.vocab_size != len(tokenizer):
resize_flag = config["train"].get("resize_model_vocab_size", True)
if not resize_flag: # 当 tokenizer 和 model 的 vocab_size 不一致，且 train_config 中指定不进行 resize 操作时，抛出一个 warning
print(f'[Vocab Warning] current model_vocab_size is `{model_config.vocab_size}`, which is not equal to tokenizer vocab_size `{len(tokenizer) }`, but `resize_model_vocab_size` is not in train_config or set as False.')
else:
raw_model.resize_token_embeddings(len(tokenizer))
print(f"Resize model vocab_size from {model_config.vocab_size} to {len(tokenizer)}.")
else:
model_config.vocab_size = len(tokenizer) # 从头开始训练
raw_model = AutoModelForCausalLM.from_config(
model_config,
trust_remote_code=True
)
if config["train"].get("train_token_embeddings_only", False): # 是否只训练 token embedding
print('[Train Token Embedding Only] Start to freeze model...')
token_embedding_layer_name = config["train"].get("token_embedding_layer_name", "")
assert token_embedding_layer_name, "[Train Token Embedding Only] Can not find `token_embedding_layer_name` in train config, or you can set `train_token_embeddings_only` to False."
from utils.model_freeze import freeze_model_exclude_token_embeddings
freeze_model_exclude_token_embeddings(
raw_model,
token_embedding_layer_name
)
print('[Train Token Embedding Only] Initialized Model.')
if config["train"].get("use_lora", False):
if hasattr(raw_model, "enable_input_require_grads"):
raw_model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
raw_model.get_input_embeddings().register_forward_hook(
make_inputs_require_grad
)
target_modules = config['train'].get('target_modules', [])
assert target_modules, 'You must specify `target_modules` since you want to use LoRA.'
peft_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
target_modules=target_modules,
inference_mode=False,
r=8,
lora_alpha=32,
lora_dropout=0.1,
)
raw_model = get_peft_model(raw_model, peft_config)
raw_model.print_trainable_parameters()
if config["train"].get("gradient_checkpointing_enable", False):
raw_model.gradient_checkpointing_enable()
trainer = Trainer(
config,
raw_model,
train_loader,
tokenizer,
accelerator
)
trainer.train()
if __name__ == "__main__":
main()