-
Notifications
You must be signed in to change notification settings - Fork 59
/
tuning_train.py
278 lines (242 loc) · 11.7 KB
/
tuning_train.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
# -*- coding:utf-8 -*-
"""
This file provides a method to tuning LLaMa model with financial data:
### Lora + int8_training ###
"""
import argparse
import os
import sys
from typing import List
import torch
import transformers
from transformers import LlamaForCausalLM, LlamaTokenizer
from datasets import load_dataset
from utils.prompter import Prompter
from peft import (
LoraConfig,
get_peft_model,
get_peft_model_state_dict,
prepare_model_for_int8_training,
set_peft_model_state_dict,
)
parser = argparse.ArgumentParser()
# model/data params
parser.add_argument("--base_model", default='', type=str, required=True, help="original pretrained llama weights")
parser.add_argument("--data_path", default='yahma/alpaca-cleaned', type=str, help="dataset of SFT.")
parser.add_argument("--output_dir", default='./lora-alpaca', type=str, help="The path of lora model.")
# training hyperparams
parser.add_argument("--batch_size", default=128, type=int, help="Batch size per GPU/CPU for training.")
parser.add_argument("--micro_batch_size", default=8, type=int, help="Batch size per process for training.")
parser.add_argument("--num_epochs", default=8, type=int, help="Total number of training epochs to perform.")
parser.add_argument("--learning_rate", default=2e-4, type=float, help="The initial learning rate for Adam.")
parser.add_argument("--cutoff_len", default=512, type=int,
help="The maximum total input sequence length after tokenization. Sequences longer "
"than this will be truncated, sequences shorter will be padded.")
parser.add_argument("--val_set_size", default=1200, type=int, help="Batch size for evaluate.")
# lora hyperparams
parser.add_argument("--lora_r", default=8, type=int, help="The number of Lora ranks.")
parser.add_argument("--lora_alpha", default=16, type=int, help="Set the alpha parameter of LORA.")
parser.add_argument("--lora_dropout", default=0.05, type=float, help="Set the dropout parameter of LORA.")
parser.add_argument("--lora_target_modules", default=["q_proj", "v_proj"], type=List[str],
help="Set the target module for the PEFT model.")
# llm hyperparams
parser.add_argument("--train_on_inputs", default=False, type=bool, help="if False, masks out inputs in loss.")
parser.add_argument("--group_by_length", default=False, type=bool,
help="faster, but produces an odd training loss curve.")
# wandb params
parser.add_argument("--wandb_project", default='llama_fin', type=str, help="The name of wandb_project.")
parser.add_argument("--wandb_run_name", default='', type=str)
parser.add_argument("--wandb_watch", default='', type=str, choices=['false', 'gradients', 'all'])
parser.add_argument("--wandb_log_model", default='', type=str, choices=['false', 'true'])
parser.add_argument("--resume_from_checkpoint", default=None, type=str,
help="Either training checkpoint or final adapter.")
parser.add_argument("--prompt_template_name", default='alpaca', type=str,
help="The prompt template to use, will default to alpaca.")
args = parser.parse_args()
def do_tuning():
if int(os.environ.get("LOCAL_RANK", 0)) == 0:
print(
"Training Alpaca-LoRA model with params:\n"
"base_model: {}\n".format(args.base_model),
"data_path: {}\n".format(args.data_path),
"output_dir: {}\n".format(args.output_dir),
"batch_size: {}\n".format(args.batch_size),
"micro_batch_size: {}\n".format(args.micro_batch_size),
"num_epochs: {}\n".format(args.num_epochs),
"learning_rate: {}\n".format(args.learning_rate),
"cutoff_len: {}\n".format(args.cutoff_len),
"val_set_size: {}\n".format(args.val_set_size),
"lora_r: {}\n".format(args.lora_r),
"lora_alpha: {}\n".format(args.lora_alpha),
"lora_dropout: {}\n".format(args.lora_dropout),
"lora_target_modules: {}\n".format(args.lora_target_modules),
"train_on_inputs: {}\n".format(args.train_on_inputs),
"group_by_length: {}\n".format(args.group_by_length),
"wandb_project: {}\n".format(args.wandb_project),
"wandb_run_name: {}\n".format(args.wandb_run_name),
"wandb_watch: {}\n".format(args.wandb_watch),
"wandb_log_model: {}\n".format(args.wandb_log_model),
"resume_from_checkpoint: {}\n".format(args.resume_from_checkpoint or None),
"prompt template: {}\n".format(args.prompt_template_name)
)
# --------------------------Check--------------------------
# Check if parameters passed or set in the environment
use_wandb = len(args.wandb_project) > 0 or (
"WANDB_PROJECT" in os.environ and len(os.environ["WANDB_PROJECT"]) > 0
)
# Only overwrite environment if wandb param passed
if len(args.wandb_project) > 0:
os.environ["WANDB_PROJECT"] = args.wandb_project
if len(args.wandb_watch) > 0:
os.environ["WANDB_WATCH"] = args.wandb_watch
if len(args.wandb_log_model) > 0:
os.environ["WANDB_LOG_MODEL"] = args.wandb_log_model
# Check if the base_model exists
assert args.base_model, "Please specify a base_model, for example: 'decapoda-research/llama-7b-hf'"
gradient_accumulation_steps = args.batch_size // args.micro_batch_size
device_map = "auto"
world_size = int(os.environ.get("WORLD_SIZE", 1))
ddp = world_size != 1
if ddp: # ddp adopts a multiprocessing approach
device_map = {"": int(os.environ.get("LOCAL_RANK") or 0)}
gradient_accumulation_steps = gradient_accumulation_steps // world_size
# --------------------------Model--------------------------
# 与全精度模型相比,以 8 位精度加载模型最多可节省 4 倍的内存
model = LlamaForCausalLM.from_pretrained(
args.base_model,
load_in_8bit=True,
torch_dtype=torch.float16,
device_map=device_map,
)
model = prepare_model_for_int8_training(model)
# using lora to tuning
config = LoraConfig(
r=args.lora_r,
lora_alpha=args.lora_alpha,
target_modules=args.lora_target_modules,
lora_dropout=args.lora_dropout,
bias="none",
task_type="CAUSAL_LM",
)
model = get_peft_model(model, config)
# Check the available weights and load
if args.resume_from_checkpoint:
# Full checkpoint
checkpoint_name = os.path.join(
args.resume_from_checkpoint, "pytorch_model.bin"
)
# only LoRA model
if not os.path.exists(checkpoint_name):
checkpoint_name = os.path.join(args.resume_from_checkpoint, "adapter_model.bin")
# So the trainer won't try loading its state
args.resume_from_checkpoint = None
# The two files above have a different name depending on how they were saved, but are actually the same.
if os.path.exists(checkpoint_name):
print("Restarting from {}".format(checkpoint_name))
adapters_weights = torch.load(checkpoint_name)
model = set_peft_model_state_dict(model, adapters_weights)
else:
print("Checkpoint {} not found".format(checkpoint_name))
# More transparent to x% of trainable parameters
model.print_trainable_parameters()
# --------------------------Tokenizer--------------------------
tokenizer = LlamaTokenizer.from_pretrained(args.base_model)
tokenizer.pad_token_id = 0 # 「unk」 different from the eos token
tokenizer.padding_side = "left" # allow batched inference
def tokenize(prompt, add_eos_token=True):
result = tokenizer(
prompt,
truncation=True,
max_length=args.cutoff_len,
padding=False,
return_tensors=None,
)
if (
result["input_ids"][-1] != tokenizer.eos_token_id
and len(result["input_ids"]) < args.cutoff_len
and add_eos_token
):
result["input_ids"].append(tokenizer.eos_token_id)
result["attention_mask"].append(1)
result["labels"] = result["input_ids"].copy()
return result
# choose prompt template
prompter = Prompter(args.prompt_template_name)
def generate_and_tokenize_prompt(data_point):
full_prompt = prompter.generate_prompt(
data_point["instruction"],
data_point["input"],
data_point["output"],
)
tokenized_full_prompt = tokenize(full_prompt)
if not args.train_on_inputs:
user_prompt = prompter.generate_prompt(data_point["instruction"], data_point["input"])
tokenized_user_prompt = tokenize(user_prompt, add_eos_token=False)
user_prompt_len = len(tokenized_user_prompt["input_ids"])
tokenized_full_prompt["labels"] = [-100] * user_prompt_len \
+ tokenized_full_prompt["labels"][user_prompt_len:] # Maybe faster
return tokenized_full_prompt
# load dataset from file(here is xx.json)
if args.data_path.endswith(".json") or args.data_path.endswith(".jsonl"):
data = load_dataset("json", data_files=args.data_path)
else:
data = load_dataset(args.data_path)
if args.val_set_size > 0:
train_val = data["train"].train_test_split(
test_size=args.val_set_size, shuffle=True, seed=2023
)
train_data = (
train_val["train"].shuffle().map(generate_and_tokenize_prompt)
)
val_data = (
train_val["test"].shuffle().map(generate_and_tokenize_prompt)
)
else:
train_data = data["train"].shuffle().map(generate_and_tokenize_prompt)
val_data = None
# --------------------------Trainer--------------------------
# 当多张显卡时,阻止 Trainer 使用自己的 DataParallelism
if not ddp and torch.cuda.device_count() > 1:
model.is_parallelizable = True
model.model_parallel = True
trainer = transformers.Trainer(
model=model,
train_dataset=train_data,
eval_dataset=val_data,
args=transformers.TrainingArguments(
per_device_train_batch_size=args.micro_batch_size,
gradient_accumulation_steps=gradient_accumulation_steps,
warmup_ratio=0.1,
num_train_epochs=args.num_epochs,
learning_rate=args.learning_rate,
fp16=True,
logging_steps=8,
optim="adamw_torch",
evaluation_strategy="steps" if args.val_set_size > 0 else "no",
save_strategy="steps",
eval_steps=32 if args.val_set_size > 0 else None,
save_steps=32,
output_dir=args.output_dir,
save_total_limit=5,
load_best_model_at_end=True if args.val_set_size > 0 else False,
ddp_find_unused_parameters=False if ddp else None,
group_by_length=args.group_by_length,
report_to="wandb" if use_wandb else None,
run_name=args.wandb_run_name if use_wandb else None,
),
data_collator=transformers.DataCollatorForSeq2Seq(
tokenizer, pad_to_multiple_of=8, return_tensors="pt", padding=True
),
)
model.config.use_cache = False
old_state_dict = model.state_dict
model.state_dict = (
lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())
).__get__(model, type(model))
if torch.__version__ >= "2" and sys.platform != "win32":
model = torch.compile(model)
trainer.train(resume_from_checkpoint=args.resume_from_checkpoint)
model.save_pretrained(args.output_dir)
print("\n 若上面出现有关于keys丢失的警告,请忽略! o(^_^)o ~")
if __name__ == "__main__":
do_tuning()