File size: 6,308 Bytes
083d82d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

print(device)


from datasets import load_dataset
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
    HfArgumentParser,
    TrainingArguments,
    pipeline,
    logging,
    LlamaTokenizerFast
)
from peft import LoraConfig, PeftModel, get_peft_model
from trl import SFTTrainer

# Le modèle que nous allons utiliser dans le Hugging Face hub
model_name = "mistral-hermes"

torch.cuda.empty_cache()

#project_directory = "~/finetuning/sigmund-spplus"

# Le nom du nouveau modèle
new_model_name = "mistral-mfs-reference"

# The output directory where the model predictions and checkpoints will be written
output_dir = "./mistral-mfs-reference"

# Tensorboard logs
tb_log_dir = "./mistral-mfs-reference/logs"

# Nombre de steps : à ajuster selon la taille du corpus et le nombre d'epochs à faire tourner.
max_steps = 500

# Les paramètres importants !!
per_device_train_batch_size = 4 #Nombre d'exemples envoyés par batch. En mettre plus pour aller plus vite.
learning_rate = 2e-5 #De préférence un taux d'apprentissage bas comme Mistral-Hermes se débrouille bien en français
max_seq_length = 4096 #C'est la fenêtre contextuelle. Elle peut être portée jusqu'à 4096 tokens (mais attention à la mémoire disponible !)
save_steps = 1000 # Sauvegarde des steps (permet de faire redémarrer l'entraînement si le fine-tuning ne fonctionne pas)
# Learning rate schedule (constant a bit better than cosine, and has advantage for analysis)
lr_scheduler_type = "linear"


#Les autres paramètres
local_rank = -1
per_device_eval_batch_size = 1
gradient_accumulation_steps = 4
max_grad_norm = 0.3
weight_decay = 0.001
lora_alpha = 16
lora_dropout = 0.1
lora_r = 64
# Group sequences into batches with same length (saves memory and speeds up training considerably)
group_by_length = True

# Activate 4-bit precision base model loading
use_4bit = True

# Activate nested quantization for 4-bit base models
use_nested_quant = False

# Compute dtype for 4-bit base models
bnb_4bit_compute_dtype = "float16"

# Quantization type (fp4 or nf4=
bnb_4bit_quant_type = "nf4"

# Number of training epochs
num_train_epochs = 1

# Enable fp16 training
fp16 = True

# Enable bf16 training
bf16 = False

# Use packing dataset creating
packing = False

# Enable gradient checkpointing
gradient_checkpointing = True

# Optimizer to use, original is paged_adamw_32bit
optim = "paged_adamw_32bit"

# Fraction of steps to do a warmup for
warmup_ratio = 0.03

# Log every X updates steps
logging_steps = 1

# Load the entire model on the GPU 0
device_map = {"": 0}

# Visualize training
report_to = "tensorboard"


#2. Import du tokenizer.
peft_config = LoraConfig(
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    r=lora_r,
    inference_mode=False,
    task_type="CAUSAL_LM",
    target_modules = ["q_proj", "v_proj"]
)

tokenizer = AutoTokenizer.from_pretrained(model_name)

# This is the fix for fp16 training
#tokenizer.padding_side = "right"
#tokenizer.pad_token = tokenizer.eos_token

#3. Préparation de la base de données

from datasets import load_dataset

def format_alpaca(sample):
    prompt = f"{sample['conversation']}"
    return prompt

# template dataset to add prompt to each sample
def template_dataset(sample):
    sample["text"] = f"{format_alpaca(sample)}{tokenizer.eos_token}"
    return sample

# Chargement du dataset.
#dataset = load_dataset("databricks/databricks-dolly-15k", split="train")
data_files = {"train": "references_mfs_corpus.json"}
dataset = load_dataset("json", data_files=data_files, split="train")

# Shuffle the dataset
dataset_shuffled = dataset.shuffle(seed=42)

# Select the first 250 rows from the shuffled dataset, comment if you want 15k
#dataset = dataset_shuffled.select(range(512))

#Transformation du dataset pour utiliser le format guanaco
dataset = dataset.map(template_dataset, remove_columns=list(dataset.features))

print(dataset[40])

#4. Import du modèle

# Load tokenizer and model with QLoRA configuration
compute_dtype = getattr(torch, bnb_4bit_compute_dtype)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=use_4bit,
    bnb_4bit_quant_type=bnb_4bit_quant_type,
    bnb_4bit_compute_dtype=compute_dtype,
    bnb_4bit_use_double_quant=use_nested_quant,
)

if compute_dtype == torch.float16 and use_4bit:
    major, _ = torch.cuda.get_device_capability()
    if major >= 8:
        print("=" * 80)
        print("Your GPU supports bfloat16, you can accelerate training with the argument --bf16")
        print("=" * 80)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map=device_map,
    quantization_config=bnb_config
)

model.config.use_cache = False
model.config.pretraining_tp = 1

#5. Fine-tuning

torch.cuda.empty_cache()

training_arguments = TrainingArguments(
    output_dir=output_dir,
    per_device_train_batch_size=per_device_train_batch_size,
    gradient_accumulation_steps=gradient_accumulation_steps,
    gradient_checkpointing=True,
    optim=optim,
    save_steps=save_steps,
    logging_steps=logging_steps,
    learning_rate=learning_rate,
    fp16=fp16,
    bf16=bf16,
    max_grad_norm=max_grad_norm,
    max_steps=max_steps,
    warmup_ratio=warmup_ratio,
    group_by_length=group_by_length,
    lr_scheduler_type=lr_scheduler_type,
    report_to="tensorboard"
)

trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    peft_config=peft_config,
    dataset_text_field="text",
    max_seq_length=max_seq_length,
    tokenizer=tokenizer,
    args=training_arguments,
    packing=packing
)

trainer.train()
#trainer.train(resume_from_checkpoint=True)

#6. Sauvegarde

model_to_save = trainer.model.module if hasattr(trainer.model, 'module') else trainer.model  # Take care of distributed/parallel training
model_to_save.save_pretrained(new_model_name)

torch.cuda.empty_cache()

from peft import AutoPeftModelForCausalLM

model = AutoPeftModelForCausalLM.from_pretrained(new_model_name, device_map="auto", torch_dtype=torch.bfloat16)
model = model.merge_and_unload()

output_merged_dir = os.path.join(new_model_name, new_model_name)
model.save_pretrained(output_merged_dir, safe_serialization=True)

tokenizer.save_pretrained(output_merged_dir)