scripts/train.py

"""
Fine-tune Qwen2.5-3B-Instruct with Unsloth (4-bit QLoRA + SFT).

Installation (CUDA GPU required):
    pip install -r modal_apps/requirements-modal.txt

Or install manually:
    pip install torch torchvision
    pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
    pip install unsloth_zoo
    pip install --no-deps trl peft accelerate bitsandbytes
    pip install datasets transformers

Run:
    python scripts/train.py
"""

import unsloth  # noqa: F401 — must import before trl/transformers/peft

from pathlib import Path

from datasets import load_dataset
from trl import SFTConfig, SFTTrainer
from unsloth import FastLanguageModel, is_bf16_supported
from unsloth.chat_templates import get_chat_template

MODEL_NAME = "Qwen/Qwen2.5-3B-Instruct"
PROJECT_ROOT = Path(__file__).resolve().parent.parent
DATA_PATH = str(PROJECT_ROOT / "data" / "train.jsonl")
OUTPUT_DIR = str(PROJECT_ROOT / "trained_model")
MAX_SEQ_LENGTH = 2048


def main() -> None:
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name=MODEL_NAME,
        max_seq_length=MAX_SEQ_LENGTH,
        dtype=None,
        load_in_4bit=True,
    )

    model = FastLanguageModel.get_peft_model(
        model,
        r=16,
        target_modules=[
            "q_proj",
            "k_proj",
            "v_proj",
            "o_proj",
            "gate_proj",
            "up_proj",
            "down_proj",
        ],
        lora_alpha=16,
        lora_dropout=0,
        bias="none",
        use_gradient_checkpointing="unsloth",
        random_state=3407,
        max_seq_length=MAX_SEQ_LENGTH,
    )

    tokenizer = get_chat_template(
        tokenizer,
        chat_template="qwen-2.5",
    )

    dataset = load_dataset("json", data_files=DATA_PATH, split="train")

    def formatting_prompts_func(examples):
        texts = [
            tokenizer.apply_chat_template(
                messages,
                tokenize=False,
                add_generation_prompt=False,
            )
            for messages in examples["messages"]
        ]
        return {"text": texts}

    dataset = dataset.map(formatting_prompts_func, batched=True)

    trainer = SFTTrainer(
        model=model,
        tokenizer=tokenizer,
        train_dataset=dataset,
        args=SFTConfig(
            output_dir=OUTPUT_DIR,
            num_train_epochs=3,
            per_device_train_batch_size=8,
            gradient_accumulation_steps=1,
            warmup_steps=10,
            learning_rate=2e-4,
            fp16=not is_bf16_supported(),
            bf16=is_bf16_supported(),
            logging_steps=10,
            optim="adamw_8bit",
            seed=3407,
            report_to="none",
            max_seq_length=MAX_SEQ_LENGTH,
            dataset_text_field="text",
            packing=False,
        ),
    )

    trainer.train()

    model.save_pretrained(OUTPUT_DIR)
    tokenizer.save_pretrained(OUTPUT_DIR)
    print(f"Model saved to {OUTPUT_DIR}")


if __name__ == "__main__":
    main()