-
Notifications
You must be signed in to change notification settings - Fork 1
/
dpo.py
243 lines (219 loc) · 8.43 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
234
235
236
237
238
239
240
241
242
243
# Adapted from https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama_2/scripts/dpo_llama2.py
# 0. imports
from dataclasses import dataclass, field
from typing import Dict, Optional
import logging
import os
import sys
import torch
import wandb
from datasets import Dataset, load_dataset
from transformers import (
AutoModelForCausalLM,
AutoTokenizer,
HfArgumentParser,
TrainingArguments,
)
from dpo_trainer import DPOExperimentalTrainer
from process_rlhf_datasets import get_hh_for_dpo
logger = logging.getLogger(__name__)
# Define and parse arguments.
@dataclass
class ScriptArguments:
"""
The arguments for the DPO training script.
"""
# training parameters
alpha: Optional[float] = field(
default=0.1, metadata={"help": "the beta parameter for DPO alpha-scaling loss"}
)
beta: Optional[float] = field(
default=0.1, metadata={"help": "the beta parameter for DPO loss"}
)
gamma: Optional[float] = field(
default=0.1, metadata={"help": "the gamma parameter for unlikelihood loss"}
)
loss_type: Optional[str] = field(
default="sigmoid",
metadata={
"help": "The type of loss objective to use. Options: ['sigmoid', 'hinge', 'ipo', 'kto_pair', 'unlikelihood']"
},
)
wandb_project: Optional[str] = field(
default="dpo", metadata={"help": "wandb project name"}
)
wandb_run_name: Optional[str] = field(
default="anthropic_hh_rlhf", metadata={"help": "wandb run name"}
)
wandb_entity: Optional[str] = field(
default="llm-calibration", metadata={"help": "wandb username or team name"}
)
model_name_or_path: Optional[str] = field(
default="gpt2", metadata={"help": "the model name"}
)
max_length: Optional[int] = field(
default=512, metadata={"help": "max length of each sample"}
)
max_prompt_length: Optional[int] = field(
default=256, metadata={"help": "max length of each sample's prompt"}
)
max_target_length: Optional[int] = field(
default=256,
metadata={
"help": "Only used for encoder decoder model. Max target of each sample's prompt"
},
)
label_pad_token_id: Optional[int] = field(
default=-100, metadata={"help": "label for non response tokens"}
)
use_flash_attn: Optional[bool] = field(
default=False,
metadata={"help": "Enables Flash attention for training."},
)
use_reentrant: Optional[bool] = field(
default=False,
metadata={"help": "Gradient Checkpointing param. Refer the related docs"},
)
# instrumentation
sanity_check: Optional[bool] = field(
default=True, metadata={"help": "only train on 1000 samples"}
)
# debug argument for distributed training
ignore_bias_buffers: Optional[bool] = field(
default=False,
metadata={
"help": "fix for DDP issues with LM bias/mask buffers - invalid scalar type,`inplace operation. See"
"https://github.com/huggingface/transformers/issues/22482#issuecomment-1595790992"
},
)
# Torch dtype used for training
torch_dtype: Optional[str] = field(
default=None,
metadata={
"help": "Torch_dtype to use when loading and training the model (e.g. 'float16,' 'bfloat16', etc."
},
)
generate_during_eval: Optional[bool] = field(
default=True,
)
dataset_name: Optional[str] = field(
default=None, metadata={"help": "the dataset for training"}
)
if __name__ == "__main__":
parser = HfArgumentParser((ScriptArguments, TrainingArguments))
script_args, training_args = parser.parse_args_into_dataclasses()
training_args.report_to = ["wandb"]
training_args.run_name = script_args.wandb_run_name
training_args.log_level = "info"
training_args.logging_first_step = True
os.environ["WANDB_PROJECT"] = script_args.wandb_project
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(training_args.log_level.upper())
logger.info(f"Training/evaluation parameters:\n{training_args}")
logger.info(f"Script args:\n{script_args}")
# 1. load a pretrained model
device_map = {"": training_args.device.index}
model_kwargs = {
"device_map": device_map,
}
if script_args.torch_dtype is not None:
assert hasattr(torch, script_args.torch_dtype)
model_kwargs["torch_dtype"] = getattr(torch, script_args.torch_dtype)
model_kwargs["attn_implementation"] = (
"flash_attention_2" if script_args.use_flash_attn else "eager"
)
model = AutoModelForCausalLM.from_pretrained(
script_args.model_name_or_path, **model_kwargs
)
if script_args.ignore_bias_buffers:
# torch distributed hack
model._ddp_params_and_buffers_to_ignore = [
name for name, buffer in model.named_buffers() if buffer.dtype == torch.bool
]
model_ref = AutoModelForCausalLM.from_pretrained(
script_args.model_name_or_path, device_map=device_map
)
tokenizer = AutoTokenizer.from_pretrained(script_args.model_name_or_path)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
# 2. Load the Anthropic Helpful-Harmless dataset
train_dataset = get_hh_for_dpo("train", sanity_check=script_args.sanity_check)
eval_and_test = get_hh_for_dpo(
"test", sanity_check=script_args.sanity_check
).train_test_split(test_size=0.5, seed=0)
eval_dataset = eval_and_test["train"]
test_dataset = eval_and_test["test"]
if script_args.sanity_check:
training_args.eval_steps = 10
wandb.init(
project=script_args.wandb_project,
name=script_args.wandb_run_name,
entity=script_args.wandb_entity,
)
# gradient ckpt
model.config.use_cache = not training_args.gradient_checkpointing
if training_args.gradient_checkpointing:
training_args.gradient_checkpointing_kwargs = {
"use_reentrant": script_args.use_reentrant
}
# 5. initialize the DPO trainer
dpo_trainer = DPOExperimentalTrainer(
model,
model_ref,
args=training_args,
alpha=script_args.alpha,
beta=script_args.beta,
gamma=script_args.gamma,
loss_type=script_args.loss_type,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
tokenizer=tokenizer,
max_length=script_args.max_length,
max_target_length=script_args.max_target_length,
max_prompt_length=script_args.max_prompt_length,
generate_during_eval=script_args.generate_during_eval,
)
# hack because DPOTrainer doesn't properly place both
# models on the same device
if dpo_trainer.is_fsdp_enabled:
prepared_model = dpo_trainer._wrap_model(
dpo_trainer.model, training=True, dataloader=None
)
if hasattr(dpo_trainer.lr_scheduler, "step"):
prepared_model, dpo_trainer.optimizer = dpo_trainer.accelerator.prepare(
prepared_model, dpo_trainer.optimizer
)
else:
(
prepared_model,
dpo_trainer.optimizer,
dpo_trainer.lr_scheduler,
) = dpo_trainer.accelerator.prepare(
prepared_model, dpo_trainer.optimizer, dpo_trainer.lr_scheduler
)
dpo_trainer.model_wrapped = prepared_model
dpo_trainer.model = prepared_model
if dpo_trainer.ref_model is not None:
dpo_trainer.ref_model = dpo_trainer.accelerator.prepare_model(
dpo_trainer.ref_model
)
dpo_trainer.accelerator.prepare_model = (
lambda model, *args, **kwargs: model
) # Monkey-patch prepare_model a no-op , since we have manually prepared the models
# 6. train
checkpoint = None
if training_args.resume_from_checkpoint is not None:
if training_args.resume_from_checkpoint in ["true", "True"]:
checkpoint = True
elif training_args.resume_from_checkpoint in ["false", "False"]:
checkpoint = False
else:
checkpoint = training_args.resume_from_checkpoint
dpo_trainer.train(resume_from_checkpoint=checkpoint)
# 7. save final model!
if not script_args.sanity_check:
dpo_trainer.save_model(training_args.output_dir)