llm-finetune

[project]
name = "llm-finetune"
version = "0.1.0"
description = "LLM fine-tuning environment"
requires-python = "==3.12.*"
dependencies = [
  "torch==2.10.0",
  "torchvision==0.25.0",
  "torchaudio==2.10.0",
  "transformers==4.57.6",
  "datasets==4.3.0",
  "accelerate==1.12.0",
  "peft==0.18.1",
  "trl==0.24.0",
  "bitsandbytes==0.49.2",
  "xformers==0.0.35; platform_system != 'Darwin'",
  "unsloth==2026.2.1; platform_system == 'Linux' and platform_machine == 'x86_64'",
  "sentencepiece==0.2.1",
  "safetensors==0.7.0",
  "ipykernel==7.2.0",
  "pytest==9.0.2",
  "ruff==0.15.2",
]

[tool.uv]
package = false

uv run python -c "import torch, transformers, peft, trl, datasets; print('torch', torch.__version__); print('transformers', transformers.__version__); print('trl', trl.__version__); print('cuda', torch.cuda.is_available())"

uv run python -c "import unsloth; print('unsloth', unsloth.__version__)"

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset

# 1) 4bit 양자화 설정
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype="bfloat16",
)

# 2) 모델 및 토크나이저 로딩
model_id = "meta-llama/Llama-3.1-8B-Instruct"
model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# 3) LoRA 어댑터 설정
lora_config = LoraConfig(r=16, lora_alpha=32, target_modules=["q_proj", "v_proj"], lora_dropout=0.05)

# 4) 데이터셋 준비
dataset = load_dataset("json", data_files="train.jsonl", split="train")

# 5) SFT 학습
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    peft_config=lora_config,
    args=SFTConfig(output_dir="./output", num_train_epochs=3, per_device_train_batch_size=4),
    tokenizer=tokenizer,
)
trainer.train()
trainer.save_model("./output/final")