File size: 12,309 Bytes
d660b02
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
import argparse
import os
from pathlib import Path

from unsloth import PatchDPOTrainer

PatchDPOTrainer()

from typing import Any, List, Literal, Optional  # noqa: E402

import torch  # noqa
from datasets import concatenate_datasets, load_dataset  # noqa: E402
from huggingface_hub import HfApi  # noqa: E402
from huggingface_hub.utils import RepositoryNotFoundError  # noqa: E402
from transformers import TextStreamer, TrainingArguments  # noqa: E402
from trl import DPOConfig, DPOTrainer, SFTTrainer  # noqa: E402
from unsloth import FastLanguageModel, is_bfloat16_supported  # noqa: E402
from unsloth.chat_templates import get_chat_template  # noqa: E402

alpaca_template = """Below is an instruction that describes a task. Write a response that appropriately completes the request.



### Instruction:

{}



### Response:

{}"""


def load_model(

    model_name: str,

    max_seq_length: int,

    load_in_4bit: bool,

    lora_rank: int,

    lora_alpha: int,

    lora_dropout: float,

    target_modules: List[str],

    chat_template: str,

) -> tuple:
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name=model_name,
        max_seq_length=max_seq_length,
        load_in_4bit=load_in_4bit,
    )

    model = FastLanguageModel.get_peft_model(
        model,
        r=lora_rank,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        target_modules=target_modules,
    )

    tokenizer = get_chat_template(
        tokenizer,
        chat_template=chat_template,
    )

    return model, tokenizer


def finetune(

    finetuning_type: Literal["sft", "dpo"],

    model_name: str,

    output_dir: str,

    dataset_huggingface_workspace: str,

    max_seq_length: int = 2048,

    load_in_4bit: bool = False,

    lora_rank: int = 32,

    lora_alpha: int = 32,

    lora_dropout: float = 0.0,

    target_modules: List[str] = ["q_proj", "k_proj", "v_proj", "up_proj", "down_proj", "o_proj", "gate_proj"],  # noqa: B006

    chat_template: str = "chatml",

    learning_rate: float = 3e-4,

    num_train_epochs: int = 3,

    per_device_train_batch_size: int = 2,

    gradient_accumulation_steps: int = 8,

    beta: float = 0.5,  # Only for DPO

    is_dummy: bool = True,

) -> tuple:
    model, tokenizer = load_model(
        model_name, max_seq_length, load_in_4bit, lora_rank, lora_alpha, lora_dropout, target_modules, chat_template
    )
    EOS_TOKEN = tokenizer.eos_token
    print(f"Setting EOS_TOKEN to {EOS_TOKEN}")  # noqa

    if is_dummy is True:
        num_train_epochs = 1
        print(f"Training in dummy mode. Setting num_train_epochs to '{num_train_epochs}'")  # noqa
        print(f"Training in dummy mode. Reducing dataset size to '400'.")  # noqa

    if finetuning_type == "sft":

        def format_samples_sft(examples):
            text = []
            for instruction, output in zip(examples["instruction"], examples["output"], strict=False):
                message = alpaca_template.format(instruction, output) + EOS_TOKEN
                text.append(message)

            return {"text": text}

        dataset1 = load_dataset(f"{dataset_huggingface_workspace}/llmtwin", split="train")
        dataset2 = load_dataset("mlabonne/FineTome-Alpaca-100k", split="train[:10000]")
        dataset = concatenate_datasets([dataset1, dataset2])
        if is_dummy:
            dataset = dataset.select(range(400))
        print(f"Loaded dataset with {len(dataset)} samples.")  # noqa

        dataset = dataset.map(format_samples_sft, batched=True, remove_columns=dataset.column_names)
        dataset = dataset.train_test_split(test_size=0.05)

        print("Training dataset example:")  # noqa
        print(dataset["train"][0])  # noqa

        trainer = SFTTrainer(
            model=model,
            tokenizer=tokenizer,
            train_dataset=dataset["train"],
            eval_dataset=dataset["test"],
            dataset_text_field="text",
            max_seq_length=max_seq_length,
            dataset_num_proc=2,
            packing=True,
            args=TrainingArguments(
                learning_rate=learning_rate,
                num_train_epochs=num_train_epochs,
                per_device_train_batch_size=per_device_train_batch_size,
                gradient_accumulation_steps=gradient_accumulation_steps,
                fp16=not is_bfloat16_supported(),
                bf16=is_bfloat16_supported(),
                logging_steps=1,
                optim="adamw_8bit",
                weight_decay=0.01,
                lr_scheduler_type="linear",
                per_device_eval_batch_size=per_device_train_batch_size,
                warmup_steps=10,
                output_dir=output_dir,
                report_to="comet_ml",
                seed=0,
            ),
        )
    elif finetuning_type == "dpo":
        PatchDPOTrainer()

        def format_samples_dpo(example):
            example["prompt"] = alpaca_template.format(example["prompt"], "")
            example["chosen"] = example["chosen"] + EOS_TOKEN
            example["rejected"] = example["rejected"] + EOS_TOKEN

            return {"prompt": example["prompt"], "chosen": example["chosen"], "rejected": example["rejected"]}

        dataset = load_dataset(f"{dataset_huggingface_workspace}/llmtwin-dpo", split="train")
        if is_dummy:
            dataset = dataset.select(range(400))
        print(f"Loaded dataset with {len(dataset)} samples.")  # noqa

        dataset = dataset.map(format_samples_dpo)
        dataset = dataset.train_test_split(test_size=0.05)

        print("Training dataset example:")  # noqa
        print(dataset["train"][0])  # noqa

        trainer = DPOTrainer(
            model=model,
            ref_model=None,
            tokenizer=tokenizer,
            beta=beta,
            train_dataset=dataset["train"],
            eval_dataset=dataset["test"],
            max_length=max_seq_length // 2,
            max_prompt_length=max_seq_length // 2,
            args=DPOConfig(
                learning_rate=learning_rate,
                num_train_epochs=num_train_epochs,
                per_device_train_batch_size=per_device_train_batch_size,
                gradient_accumulation_steps=gradient_accumulation_steps,
                fp16=not is_bfloat16_supported(),
                bf16=is_bfloat16_supported(),
                optim="adamw_8bit",
                weight_decay=0.01,
                lr_scheduler_type="linear",
                per_device_eval_batch_size=per_device_train_batch_size,
                warmup_steps=10,
                output_dir=output_dir,
                eval_steps=0.2,
                logging_steps=1,
                report_to="comet_ml",
                seed=0,
            ),
        )
    else:
        raise ValueError("Invalid finetuning_type. Choose 'sft' or 'dpo'.")

    trainer.train()

    return model, tokenizer


def inference(

    model: Any,

    tokenizer: Any,

    prompt: str = "Write a paragraph to introduce supervised fine-tuning.",

    max_new_tokens: int = 256,

) -> None:
    model = FastLanguageModel.for_inference(model)
    message = alpaca_template.format(prompt, "")
    inputs = tokenizer([message], return_tensors="pt").to("cuda")

    text_streamer = TextStreamer(tokenizer)
    _ = model.generate(**inputs, streamer=text_streamer, max_new_tokens=max_new_tokens, use_cache=True)


def save_model(model: Any, tokenizer: Any, output_dir: str, push_to_hub: bool = False, repo_id: Optional[str] = None):
    model.save_pretrained_merged(output_dir, tokenizer, save_method="merged_16bit")

    if push_to_hub and repo_id:
        print(f"Saving model to '{repo_id}'")  # noqa
        model.push_to_hub_merged(repo_id, tokenizer, save_method="merged_16bit")


def check_if_huggingface_model_exists(model_id: str, default_value: str = "mlabonne/TwinLlama-3.1-8B") -> str:
    api = HfApi()

    try:
        api.model_info(model_id)
    except RepositoryNotFoundError:
        print(f"Model '{model_id}' does not exist.")  # noqa
        model_id = default_value
        print(f"Defaulting to '{model_id}'")  # noqa
        print("Train your own 'TwinLlama-3.1-8B' to avoid this behavior.")  # noqa

    return model_id


if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    parser.add_argument("--num_train_epochs", type=int, default=3)
    parser.add_argument("--per_device_train_batch_size", type=int, default=2)
    parser.add_argument("--learning_rate", type=float, default=3e-4)
    parser.add_argument("--dataset_huggingface_workspace", type=str, default="mlabonne")
    parser.add_argument("--model_output_huggingface_workspace", type=str, default="mlabonne")
    parser.add_argument("--is_dummy", type=bool, default=False, help="Flag to reduce the dataset size for testing")
    parser.add_argument(
        "--finetuning_type",
        type=str,
        choices=["sft", "dpo"],
        default="sft",
        help="Parameter to choose the finetuning stage.",
    )

    parser.add_argument("--output_data_dir", type=str, default=os.environ["SM_OUTPUT_DATA_DIR"])
    parser.add_argument("--model_dir", type=str, default=os.environ["SM_MODEL_DIR"])
    parser.add_argument("--n_gpus", type=str, default=os.environ["SM_NUM_GPUS"])

    args = parser.parse_args()

    print(f"Num training epochs: '{args.num_train_epochs}'")  # noqa
    print(f"Per device train batch size: '{args.per_device_train_batch_size}'")  # noqa
    print(f"Learning rate: {args.learning_rate}")  # noqa
    print(f"Datasets will be loaded from Hugging Face workspace: '{args.dataset_huggingface_workspace}'")  # noqa
    print(f"Models will be saved to Hugging Face workspace: '{args.model_output_huggingface_workspace}'")  # noqa
    print(f"Training in dummy mode? '{args.is_dummy}'")  # noqa
    print(f"Finetuning type: '{args.finetuning_type}'")  # noqa

    print(f"Output data dir: '{args.output_data_dir}'")  # noqa
    print(f"Model dir: '{args.model_dir}'")  # noqa
    print(f"Number of GPUs: '{args.n_gpus}'")  # noqa

    if args.finetuning_type == "sft":
        print("Starting SFT training...")  # noqa
        base_model_name = "meta-llama/Meta-Llama-3.1-8B"
        print(f"Training from base model '{base_model_name}'")  # noqa

        output_dir_sft = Path(args.model_dir) / "output_sft"
        model, tokenizer = finetune(
            finetuning_type="sft",
            model_name=base_model_name,
            output_dir=str(output_dir_sft),
            dataset_huggingface_workspace=args.dataset_huggingface_workspace,
            num_train_epochs=args.num_train_epochs,
            per_device_train_batch_size=args.per_device_train_batch_size,
            learning_rate=args.learning_rate,
        )
        inference(model, tokenizer)

        sft_output_model_repo_id = f"{args.model_output_huggingface_workspace}/TwinLlama-3.1-8B"
        save_model(model, tokenizer, "model_sft", push_to_hub=True, repo_id=sft_output_model_repo_id)
    elif args.finetuning_type == "dpo":
        print("Starting DPO training...")  # noqa

        sft_base_model_repo_id = f"{args.model_output_huggingface_workspace}/TwinLlama-3.1-8B"
        sft_base_model_repo_id = check_if_huggingface_model_exists(sft_base_model_repo_id)
        print(f"Training from base model '{sft_base_model_repo_id}'")  # noqa

        output_dir_dpo = Path(args.model_dir) / "output_dpo"
        model, tokenizer = finetune(
            finetuning_type="dpo",
            model_name=sft_base_model_repo_id,
            output_dir=str(output_dir_dpo),
            dataset_huggingface_workspace=args.dataset_huggingface_workspace,
            num_train_epochs=1,
            per_device_train_batch_size=args.per_device_train_batch_size,
            learning_rate=2e-6,
            is_dummy=args.is_dummy,
        )
        inference(model, tokenizer)

        dpo_output_model_repo_id = f"{args.model_output_huggingface_workspace}/TwinLlama-3.1-8B-DPO"
        save_model(model, tokenizer, "model_dpo", push_to_hub=True, repo_id=dpo_output_model_repo_id)