src/training/sft/train_sft_hydra.py

#!/usr/bin/env python3
"""
SFT training script with Hydra for LoRA (resume supported)
"""

import os
import sys
import logging
from pathlib import Path

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

import hydra
from omegaconf import DictConfig, OmegaConf
from datasets import Dataset
from peft import (
    LoraConfig,
    get_peft_model,
    PeftModel,
)
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
)
from trl import SFTTrainer, SFTConfig
from huggingface_hub import login

# dataset utils
from src.dataset.sft import (
    load_csv_dataset_sft,
    map_problem_description_to_conversation_sft,
)

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


# ------------------------------------------------------------
# Dataset loader
# ------------------------------------------------------------
def load_train_dataset(cfg: DictConfig) -> Dataset:
    raw_dataset = load_csv_dataset_sft(
        cfg.file_path, map_problem_description_to_conversation_sft
    )
    raw_dataset = raw_dataset.shuffle(seed=cfg.seed)
    train_dataset = raw_dataset.select(range(min(cfg.max_rows, len(raw_dataset))))
    logger.info("Train rows: %d", len(train_dataset))
    return train_dataset


# ------------------------------------------------------------
# Create model + LoRA
# ------------------------------------------------------------
def create_lora_model(cfg, resume_path=None):
    """
    RULE:
    - If resume_path provided: load base model then load LoRA adapter
    - Else: load base model then attach new LoRA
    """

    base_model_id = cfg.model_id

    logger.info(f"Loading base model: {base_model_id}")
    base_model = AutoModelForCausalLM.from_pretrained(
        base_model_id,
        device_map=cfg.device_map,
    )

    if resume_path:
        logger.info(f"Resume from LoRA adapter: {resume_path}")
        model = PeftModel.from_pretrained(base_model, resume_path)
        return model

    # Create new LoRA
    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,
    )

    model = get_peft_model(base_model, lora_cfg)
    logger.info("New LoRA model created")

    return model


# ------------------------------------------------------------
# TrainingConfig
# ------------------------------------------------------------
def build_sft_config(cfg, output_dir):
    return SFTConfig(
        output_dir=output_dir,
        learning_rate=cfg.learning_rate,
        weight_decay=cfg.weight_decay,
        warmup_ratio=cfg.warmup_ratio,
        gradient_accumulation_steps=cfg.gradient_accumulation_steps,
        per_device_train_batch_size=cfg.per_device_train_batch_size,
        num_train_epochs=cfg.num_train_epochs,
        max_length=cfg.max_length,
        bf16=cfg.bf16,
        fp16=cfg.fp16,
        logging_steps=cfg.logging_steps,
        save_strategy=cfg.save_strategy,
        save_steps=cfg.save_steps,
        report_to=cfg.report_to,
        lr_scheduler_type=cfg.lr_scheduler_type,
        optim=cfg.optim,
        remove_unused_columns=cfg.remove_unused_columns,
    )


# ------------------------------------------------------------
# Trainer
# ------------------------------------------------------------
def create_trainer(model, tokenizer, train_dataset, training_args):
    return SFTTrainer(
        model=model,
        train_dataset=train_dataset,
        args=training_args,
        tokenizer=tokenizer,
    )


# ------------------------------------------------------------
# Train & Save (LoRA ONLY)
# ------------------------------------------------------------
def train_and_save(trainer, output_dir, tokenizer, hf_repo_id=None):
    logger.info("Start training...")
    trainer.train()
    logger.info("Training finished")

    # SAVE ONLY LORA ADAPTER
    trainer.model.save_pretrained(output_dir)
    tokenizer.save_pretrained(output_dir)

    logger.info(f"Saved LoRA adapter to: {output_dir}")

    if hf_repo_id:
        logger.info(f"Pushing adapter to HF Hub: {hf_repo_id}")
        trainer.model.push_to_hub(hf_repo_id)
        tokenizer.push_to_hub(hf_repo_id)


# ------------------------------------------------------------
# MAIN
# ------------------------------------------------------------
@hydra.main(version_base=None, config_path="../../config/sft", config_name="config")
def main(cfg: DictConfig):
    print("Loaded config:")
    print(OmegaConf.to_yaml(cfg))

    # Login HF (optional)
    if cfg.get("hf_token", None):
        login(cfg.hf_token)
        logger.info("Logged into HF")

    # Check dataset
    if not Path(cfg.dataset.file_path).exists():
        logger.error(f"Dataset not found: {cfg.dataset.file_path}")
        return

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

    # Load dataset
    train_dataset = load_train_dataset(cfg.dataset)

    # Load tokenizer
    tokenizer = AutoTokenizer.from_pretrained(cfg.model.model_id)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    # Create model (resume if available)
    resume = cfg.get("resume_from", None)
    model = create_lora_model(cfg.model, resume)

    # Training configuration
    training_args = build_sft_config(cfg.training, cfg.output_dir)

    # Trainer
    trainer = create_trainer(model, tokenizer, train_dataset, training_args)

    # Train & Save
    train_and_save(
        trainer=trainer,
        output_dir=cfg.output_dir,
        tokenizer=tokenizer,
        hf_repo_id=cfg.get("hf_repo_id", None),
    )


if __name__ == "__main__":
    main()