File size: 6,876 Bytes
00db46c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""

GRPO training script for arithmetic countdown problems using Hydra configuration.

After training, the model is automatically pushed to HuggingFace Hub.

"""

import logging
import os
from collections.abc import Callable
from pathlib import Path
import sys

# ---------------- FIX IMPORT WHEN USING HYDRA ----------------
FILE_DIR = os.path.dirname(os.path.abspath(__file__))  # src/training/grpo
PROJECT_ROOT = os.path.abspath(os.path.join(FILE_DIR, "../../../"))
sys.path.insert(0, PROJECT_ROOT)

import hydra
from datasets import Dataset
from omegaconf import DictConfig, OmegaConf
from transformers import AutoModelForCausalLM, PreTrainedModel, AutoTokenizer
from huggingface_hub import HfApi, login, create_repo
from peft import LoraConfig, PeftModel, get_peft_model
from trl import GRPOConfig, GRPOTrainer

from src.dataset import load_csv_dataset_grpo
from src.dataset.grpo import map_problem_description_to_conversation_grpo
from src.utils.rewards import mathematical_correctness_reward_function

# Logging
logging.basicConfig(
    level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s"
)
logger = logging.getLogger("grpo_training")


# -------------------------------------------------------------
# DATASET
# -------------------------------------------------------------
def load_train_dataset(cfg: DictConfig) -> Dataset:
    raw_dataset: Dataset = load_csv_dataset_grpo(
        cfg.file_path, cfg.split, map_problem_description_to_conversation_grpo
    )
    raw_dataset = raw_dataset.shuffle(seed=cfg.seed)
    return raw_dataset.select(range(min(cfg.max_rows, len(raw_dataset))))


# -------------------------------------------------------------
# MODEL (LoRA + resume)
# -------------------------------------------------------------
def create_lora_model(cfg: DictConfig, resume_from_checkpoint: str | None = None) -> PreTrainedModel:
    model = AutoModelForCausalLM.from_pretrained(cfg.model_id, device_map=cfg.device_map)

    if resume_from_checkpoint and Path(resume_from_checkpoint).exists():
        logger.info("Loading existing LoRA adapter and merging: %s", resume_from_checkpoint)
        model = PeftModel.from_pretrained(model, resume_from_checkpoint)
        model = model.merge_and_unload()

    lora_cfg = LoraConfig(
        r=cfg.lora.r,
        lora_alpha=cfg.lora.lora_alpha,
        target_modules=OmegaConf.to_container(cfg.lora.target_modules),
        lora_dropout=cfg.lora.lora_dropout,
        bias=cfg.lora.bias,
        task_type=cfg.lora.task_type,
    )
    return get_peft_model(model, lora_cfg)


# -------------------------------------------------------------
# GRPO CONFIG
# -------------------------------------------------------------
def create_grpo_config(cfg: DictConfig, output_dir: str) -> GRPOConfig:
    return GRPOConfig(
        output_dir=output_dir,
        learning_rate=cfg.learning_rate,
        warmup_ratio=cfg.warmup_ratio,
        weight_decay=cfg.weight_decay,
        lr_scheduler_type=cfg.lr_scheduler_type,
        optim=cfg.optim,
        remove_unused_columns=cfg.remove_unused_columns,
        gradient_accumulation_steps=cfg.gradient_accumulation_steps,
        num_train_epochs=cfg.num_train_epochs,
        bf16=cfg.bf16,
        per_device_train_batch_size=cfg.per_device_train_batch_size,
        temperature=cfg.temperature,
        max_completion_length=cfg.max_completion_length,
        num_generations=cfg.num_generations,
        max_prompt_length=cfg.max_prompt_length,
        report_to=cfg.report_to,
        logging_steps=cfg.logging_steps,
        save_strategy=cfg.save_strategy,
        save_steps=cfg.save_steps,
    )


# -------------------------------------------------------------
# TRAINER
# -------------------------------------------------------------
def create_trainer(model, train_dataset, args):
    reward_funcs = [mathematical_correctness_reward_function]
    return GRPOTrainer(
        model=model,
        reward_funcs=reward_funcs,
        args=args,
        train_dataset=train_dataset,
    )


# -------------------------------------------------------------
# TRAIN + SAVE
# -------------------------------------------------------------
def train_and_save(trainer, output_dir, resume_from_checkpoint=None, save_before_training=True):
    if save_before_training:
        trainer.save_model(output_dir)

    trainer.train(resume_from_checkpoint=resume_from_checkpoint)
    trainer.save_model(output_dir)

    logger.info("Training completed.")
    logger.info("Saved final model to: %s", output_dir)


# -------------------------------------------------------------
# PUSH TO HUGGINGFACE HUB
# -------------------------------------------------------------
def push_to_huggingface(output_dir: str, repo_id: str, model_id: str):
    logger.info("Pushing model to HuggingFace Hub...")

    # Login must be done BEFORE training
    api = HfApi()

    # Create repo if not exists
    try:
        api.create_repo(repo_id, exist_ok=True)
    except:
        pass

    # Load tokenizer (important!)
    tokenizer = AutoTokenizer.from_pretrained(model_id)

    # Push
    api.upload_folder(
        folder_path=output_dir,
        repo_id=repo_id,
        commit_message="Upload GRPO fine-tuned model",
    )

    tokenizer.push_to_hub(repo_id)

    logger.info("Upload complete! HF repo: https://huggingface.co/%s", repo_id)


# -------------------------------------------------------------
# MAIN
# -------------------------------------------------------------
@hydra.main(version_base=None, config_path="../../config/grpo", config_name="config")
def main(cfg: DictConfig):
    logger.info("Configuration:\n%s", OmegaConf.to_yaml(cfg))

    if not Path(cfg.dataset.file_path).exists():
        logger.error("Dataset CSV file does not exist: %s", cfg.dataset.file_path)
        return

    os.makedirs(cfg.output_dir, exist_ok=True)

    # Load dataset
    train_dataset = load_train_dataset(cfg.dataset)

    # Model
    resume_sft = cfg.get("resume_from_checkpoint_sft", None)
    model = create_lora_model(cfg.model, resume_sft)

    # Trainer
    training_args = create_grpo_config(cfg.training, cfg.output_dir)
    trainer = create_trainer(model, train_dataset, training_args)

    # Train
    train_and_save(
        trainer,
        cfg.output_dir,
        resume_from_checkpoint=cfg.resume_from_checkpoint_grpo,
        save_before_training=cfg.save_before_training,
    )

    # Push to HF
    if cfg.get("push_to_hub", False):
        push_to_huggingface(
            output_dir=cfg.output_dir,
            repo_id=cfg.hf_repo_id,
            model_id=cfg.model.model_id,
        )


if __name__ == "__main__":
    main()