-
Notifications
You must be signed in to change notification settings - Fork 2
/
Finetune_SFTTrainer_withLoRA_OneChatbotGPT2Vi.py
274 lines (233 loc) · 9.55 KB
/
Finetune_SFTTrainer_withLoRA_OneChatbotGPT2Vi.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
# -*- coding: utf-8 -*-
# Author: Mr.Jack _ www.BICweb.vn
# Date: 26 May 2024
# https://huggingface.co/docs/trl/en/sft_trainer
# https://huggingface.co/docs/peft/en/developer_guides/model_merging
import os, torch
from trl import SFTTrainer
import transformers
from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import Trainer, TrainingArguments
os.environ["TOKENIZERS_PARALLELISM"] = "False"
RANDOM_SEED = 3407 # 3407 , 42
# device = torch.device('mps')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
MODEL_NAME = 'OneChatbotGPT2Vi'
# MODEL_NAME = './test_trainer/checkpoint-10'
print("MODEL_NAME:",MODEL_NAME)
# Step 1: Pretrained loading
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model.config.use_cache = False
# pad_token_id=2
if tokenizer.pad_token_id is None:
tokenizer.add_special_tokens({'pad_token': '[PAD]'})
print("Add new pad_token: [PAD]")
# text = 'Question: Xin chào\n Answer: Công ty BICweb kính chào quý khách!.'
# text = "Question: Xin chào Answer: Dạ, em chào anh, dạo này anh có khỏe không ạ!."
text = "Question: Xin chào Answer: Dạ, em kính chào quý anh ạ!."
print("text:",text)
data = [{"text": text}]
# data = [{"input_ids": tokenizer.encode(text=text, add_special_tokens=True, return_tensors='pt')}]
from datasets import Dataset
dataset = Dataset.from_list(data)
# dataset.set_format("torch")
# print(dataset)
# print(dataset[0])
EPOCHS = 80
LEARNING_RATE = 3e-4
OUTPUT_DIR = "test_trainer"
# print(model)
from peft import LoraConfig, get_peft_model
peft_config = LoraConfig(
init_lora_weights="gaussian",
r=16, # 16, 32, 64, 128, 256
lora_alpha=32, # 32, 64, 128, 256
lora_dropout=0.05,
bias="none",
task_type="CAUSAL_LM",
fan_in_fan_out=True,
target_modules=[
"attn.c_attn",
"attn.c_proj",
"mlp.c_fc",
"mlp.c_proj",
]
)
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()
print("\n")
# RANK: r=16 ; epochs=50 ; checkpoint file: ~10MB ; target modules: Conv1D()
# trainable params: 589,824 || all params: 125,029,632 || trainable%: 0.4717473694555863
# (Ok) RANK: r=16 ; lora_alpha=32 ; epochs=80 ; checkpoint file: ~32MB ; adapter_model.safetensors: ~9.4MB; with target_modules: ["attn.c_attn", "attn.c_proj", "mlp.c_fc", "mlp.c_proj", ]
# trainable params: 2,359,296 || all params: 126,799,104 || trainable%: 1.8606566809809635
# (Ok) RANK: r=32 ; lora_alpha=32 ; epochs=80 ; checkpoint file: ~60MB ; adapter_model.safetensors: ~18.9MB; with target_modules: ["attn.c_attn", "attn.c_proj", "mlp.c_fc", "mlp.c_proj", ]
# trainable params: 4,718,592 || all params: 129,158,400 || trainable%: 3.653337297458005
# (Ok) RANK: r=64 ; epochs=100 ; checkpoint file: ~117MB ; with target_modules: ["attn.c_attn", "attn.c_proj", "mlp.c_fc", "mlp.c_proj", ]
# trainable params: 9,437,184 || all params: 133,876,992 || trainable%: 7.049145532041831
# (Ok) RANK: r=128 ; epochs=100 ; checkpoint file: ~60MB ; target modules: Conv1D()
# trainable params: 4,718,592 || all params: 129,158,400 || trainable%: 3.653337297458005
# (Ok) RANK: r=64 ; epochs=150 ; checkpoint file: ~117MB ; adapter_model.safetensors: ~37.8MB; with target_modules: ["attn.c_attn", "attn.c_proj", "mlp.c_fc", "mlp.c_proj", ]
# trainable params: 9,437,184 || all params: 133,876,992 || trainable%: 7.049145532041831
# (Ok) RANK: r=128 ; epochs=50 ; checkpoint file: ~230MB ; adapter_model.safetensors: ~75.5MB; with target_modules: ["attn.c_attn", "attn.c_proj", "mlp.c_fc", "mlp.c_proj", ]
# trainable params: 18,874,368 || all params: 143,314,176 || trainable%: 13.169923957836522
# (Ok) RANK: r=256 ; epochs=50 ; checkpoint file: ~456MB ; with target_modules: ["attn.c_attn", "attn.c_proj", "mlp.c_fc", "mlp.c_proj", ]
# trainable params: 37,748,736 || all params: 162,188,544 || trainable%: 23.27460070176103
args_config = TrainingArguments(
num_train_epochs=EPOCHS,
learning_rate=LEARNING_RATE,
logging_steps=5, # 1, 5, 10
output_dir=OUTPUT_DIR,
seed=RANDOM_SEED, #42,
optim='adamw_torch',
lr_scheduler_type='constant_with_warmup',
gradient_accumulation_steps=10,
# per_device_train_batch_size=2,
warmup_steps=0,
weight_decay=0.01,
overwrite_output_dir=True,
save_steps=EPOCHS,
save_strategy= 'steps', # 'steps' 'epoch'
save_total_limit=1,
use_cpu=True, # default = False
)
trainer = SFTTrainer(
model=model,
train_dataset=dataset,
dataset_text_field="text", # "input_ids",
max_seq_length=256,
# tokenizer=tokenizer,
args=args_config,
peft_config=peft_config,
)
trainer.train()
# Generate responses to new questions
model.eval()
# print(model)
def generate_answer(question):
# Encode the question using the tokenizer
input_ids = tokenizer.encode(question, add_special_tokens=False, return_tensors='pt').to(device)
# Generate the answer using the model
sample_output = model.generate(input_ids, pad_token_id=2, eos_token_id=50256, max_length=256, do_sample=True, top_k=50, top_p=1.0, temperature=0.6).to(device)
# Decode the generated answer using the tokenizer
answer = tokenizer.decode(sample_output[0], skip_special_tokens=True)
sentences = answer.split('.')
return sentences[0]
# # Example usage
question = 'Question: Xin chào'
response = generate_answer(question)
print(f"\n{response}\n")
"""
PRETRAINED_MODEL_NAME: gpt2
GPT2LMHeadModel(
(transformer): GPT2Model(
(wte): Embedding(50257, 768)
(wpe): Embedding(1024, 768)
(drop): Dropout(p=0.1, inplace=False)
(h): ModuleList(
(0-11): 12 x GPT2Block(
(ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
(attn): GPT2Attention(
(c_attn): Conv1D()
(c_proj): Conv1D()
(attn_dropout): Dropout(p=0.1, inplace=False)
(resid_dropout): Dropout(p=0.1, inplace=False)
)
(ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
(mlp): GPT2MLP(
(c_fc): Conv1D()
(c_proj): Conv1D()
(act): NewGELUActivation()
(dropout): Dropout(p=0.1, inplace=False)
)
)
)
(ln_f): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
)
(lm_head): Linear(in_features=768, out_features=50257, bias=False)
)
PRETRAINED_MODEL_NAME: OneChatbotGPT2Vi
PeftModelForCausalLM(
(base_model): LoraModel(
(model): GPT2LMHeadModel(
(transformer): GPT2Model(
(wte): Embedding(50257, 768)
(wpe): Embedding(1024, 768)
(drop): Dropout(p=0.1, inplace=False)
(h): ModuleList(
(0-11): 12 x GPT2Block(
(ln_1): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
(attn): GPT2Attention(
(c_attn): lora.Linear(
(base_layer): Conv1D()
(lora_dropout): ModuleDict(
(default): Dropout(p=0.05, inplace=False)
)
(lora_A): ModuleDict(
(default): Linear(in_features=768, out_features=16, bias=False)
)
(lora_B): ModuleDict(
(default): Linear(in_features=16, out_features=2304, bias=False)
)
(lora_embedding_A): ParameterDict()
(lora_embedding_B): ParameterDict()
)
(c_proj): lora.Linear(
(base_layer): Conv1D()
(lora_dropout): ModuleDict(
(default): Dropout(p=0.05, inplace=False)
)
(lora_A): ModuleDict(
(default): Linear(in_features=768, out_features=16, bias=False)
)
(lora_B): ModuleDict(
(default): Linear(in_features=16, out_features=768, bias=False)
)
(lora_embedding_A): ParameterDict()
(lora_embedding_B): ParameterDict()
)
(attn_dropout): Dropout(p=0.1, inplace=False)
(resid_dropout): Dropout(p=0.1, inplace=False)
)
(ln_2): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
(mlp): GPT2MLP(
(c_fc): lora.Linear(
(base_layer): Conv1D()
(lora_dropout): ModuleDict(
(default): Dropout(p=0.05, inplace=False)
)
(lora_A): ModuleDict(
(default): Linear(in_features=768, out_features=16, bias=False)
)
(lora_B): ModuleDict(
(default): Linear(in_features=16, out_features=3072, bias=False)
)
(lora_embedding_A): ParameterDict()
(lora_embedding_B): ParameterDict()
)
(c_proj): lora.Linear(
(base_layer): Conv1D()
(lora_dropout): ModuleDict(
(default): Dropout(p=0.05, inplace=False)
)
(lora_A): ModuleDict(
(default): Linear(in_features=3072, out_features=16, bias=False)
)
(lora_B): ModuleDict(
(default): Linear(in_features=16, out_features=768, bias=False)
)
(lora_embedding_A): ParameterDict()
(lora_embedding_B): ParameterDict()
)
(act): NewGELUActivation()
(dropout): Dropout(p=0.1, inplace=False)
)
)
)
(ln_f): LayerNorm((768,), eps=1e-05, elementwise_affine=True)
)
(lm_head): Linear(in_features=768, out_features=50257, bias=False)
)
)
)
"""