-
Notifications
You must be signed in to change notification settings - Fork 3
/
dpo.py
233 lines (199 loc) · 6.67 KB
/
dpo.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
import os
from dataclasses import dataclass
from typing import List, Optional
import math
from datasets import load_dataset
from trl import DPOTrainer, SFTTrainer
from transformers import TrainingArguments, AddedToken, BitsAndBytesConfig
import fire
from unsloth import FastLanguageModel, PatchDPOTrainer
import torch
@dataclass
class FineTuneConfig:
model_name: str
sft_lora_path: str
eval_steps: int
save_steps: int
save_total_limit: int
output_dir: str
load_in_4bit: bool
max_seq_length: int
max_length: int
max_prompt_length: int
max_target_length: int
micro_batch_size: int
gradient_accumulation_steps: int
learning_rate: float
lora_dropout: float
lora_r: int
lora_alpha: int
lora_modules_to_save: List[str]
group_by_length: bool
gradient_checkpointing: bool
train_dataset: str
eval_dataset: str
add_tokens: List[str]
warmup_steps: Optional[int]
warmup_ratio: Optional[float]
num_train_epochs: int
seed: int
hub_strategy: Optional[str] = None
hub_model_id: Optional[str] = None
def build_trainer(config: FineTuneConfig):
assert (
config.sft_lora_path is None
), "Not supported yet. Please set model_name to your SFT model."
if config.add_tokens:
tokenizer.add_tokens(
[
AddedToken(token, rstrip=False, lstrip=False, normalized=False)
for token in config.add_tokens
]
)
# Don't think this affects tokenization from HF (tested on a lot of data), but better safe than sorry.
tokenizer.add_special_tokens({"additional_special_tokens": config.add_tokens})
train_dataset = load_dataset("json", data_files=config.train_dataset)[
"train"
].shuffle(seed=config.seed)
eval_dataset = load_dataset("json", data_files=config.eval_dataset)[
"train"
].shuffle(seed=config.seed)
dtype = None
base_model, tokenizer = FastLanguageModel.from_pretrained(
model_name=config.model_name,
max_seq_length=config.max_seq_length,
dtype=dtype,
load_in_4bit=config.load_in_4bit,
# tokens=os.environ.get("HUGGING_FACE_HUB_TOKEN", None),
)
base_model.config.use_cache = False
dpo_model = FastLanguageModel.get_peft_model(
base_model,
r=config.lora_r, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
lora_alpha=config.lora_alpha,
target_modules=[
"q_proj",
"k_proj",
"v_proj",
"o_proj",
"gate_proj",
"up_proj",
"down_proj",
],
lora_dropout=config.lora_dropout, # Currently only supports dropout = 0
bias="none", # Currently only supports bias = "none"
use_gradient_checkpointing=config.gradient_checkpointing,
random_state=config.seed,
max_seq_length=config.max_seq_length,
modules_to_save=config.lora_modules_to_save,
# adapter_name="_train",
)
report_to = None
if os.environ.get("WANDB_API_KEY", None):
report_to = "wandb"
bf16 = torch.cuda.is_bf16_supported()
dpo_trainer = DPOTrainer(
model=dpo_model,
ref_model=None,
args=TrainingArguments(
report_to=report_to,
save_strategy="steps",
save_steps=config.save_steps,
save_total_limit=config.save_total_limit,
evaluation_strategy="steps" if config.eval_steps > 0 else "no",
eval_steps=config.eval_steps,
fp16_full_eval=True,
per_device_eval_batch_size=config.micro_batch_size,
per_device_train_batch_size=config.micro_batch_size,
gradient_accumulation_steps=config.gradient_accumulation_steps,
warmup_steps=config.warmup_steps,
warmup_ratio=config.warmup_ratio,
num_train_epochs=config.num_train_epochs,
learning_rate=config.learning_rate,
fp16=not bf16,
bf16=bf16,
logging_steps=1,
optim="paged_adamw_8bit",
weight_decay=0.0,
lr_scheduler_type="linear",
seed=42,
output_dir=config.output_dir,
hub_strategy=config.hub_strategy,
hub_model_id=config.hub_model_id,
),
beta=0.1,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
tokenizer=tokenizer,
max_length=config.max_length,
max_prompt_length=config.max_prompt_length,
max_target_length=config.max_target_length,
)
return dpo_trainer
def main(
train_dataset: str,
eval_dataset: str,
model_name: str,
eval_steps: int,
save_steps: int,
save_total_limit: int,
output_dir: str,
load_in_4bit: bool = True,
max_seq_length: int = 1024,
max_length: int = 1024,
max_prompt_length: int = 1024,
max_target_length: int = 1024,
micro_batch_size: int = 1,
gradient_accumulation_steps: int = 4,
learning_rate: float = 2e-4,
lora_dropout: float = 0,
lora_r: int = 8,
lora_alpha: int = 16,
lora_modules_to_save: List[str] = [],
group_by_length: bool = False,
gradient_checkpointing: bool = True,
add_tokens: List[str] = [],
warmup_steps: int = 0,
warmup_ratio: float = 0.0,
num_train_epochs: int = 1,
hub_strategy: Optional[str] = None,
hub_model_id: Optional[str] = None,
sft_lora_path: Optional[str] = None,
seed: int = 3407,
):
PatchDPOTrainer()
config = FineTuneConfig(
train_dataset=train_dataset,
eval_dataset=eval_dataset,
model_name=model_name,
sft_lora_path=sft_lora_path,
eval_steps=eval_steps,
save_steps=save_steps,
save_total_limit=save_total_limit,
output_dir=output_dir,
load_in_4bit=load_in_4bit,
max_seq_length=max_seq_length,
max_length=max_length,
max_prompt_length=max_prompt_length,
max_target_length=max_target_length,
micro_batch_size=micro_batch_size,
gradient_accumulation_steps=gradient_accumulation_steps,
learning_rate=learning_rate,
lora_dropout=lora_dropout,
lora_r=lora_r,
lora_alpha=lora_alpha,
lora_modules_to_save=lora_modules_to_save,
group_by_length=group_by_length,
gradient_checkpointing=gradient_checkpointing,
add_tokens=add_tokens,
warmup_steps=warmup_steps,
warmup_ratio=warmup_ratio,
num_train_epochs=num_train_epochs,
hub_strategy=hub_strategy,
hub_model_id=hub_model_id,
seed=seed,
)
trainer = build_trainer(config)
trainer_stats = trainer.train()
if __name__ == "__main__":
fire.Fire(main)