import datasets
import torch
from peft import LoraConfig, TaskType, get_peft_model
from peft.peft_model import PeftModel
from transformers import LlamaForCausalLM as ModelCls
from transformers import Trainer, TrainingArguments

# 讀取 Model
model_name = "TheBloke/Llama-2-7B-Chat-fp16"
model: ModelCls = ModelCls.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

# 讀取 Peft Model
peft_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=8,
    lora_alpha=32,
    lora_dropout=0.1,
)
model: PeftModel = get_peft_model(model, peft_config)
model.print_trainable_parameters()


# 讀取資料集
data_files = {
    "train": "data/train.tokens.json.gz",
    "dev": "data/dev.tokens.json.gz",
}

dataset = datasets.load_dataset(
    "json",
    data_files=data_files,
    cache_dir="cache",
)


# 設定訓練參數
output_dir = "models/Llama-7B-TwAddr-LoRA"
train_args = TrainingArguments(
    output_dir,
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    eval_accumulation_steps=2,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    learning_rate=5e-4,
    save_total_limit=3,
    num_train_epochs=5,
    load_best_model_at_end=True,
    bf16=True,
)


# 開始訓練模型
trainer = Trainer(
    model=model,
    args=train_args,
    train_dataset=dataset["train"],
    eval_dataset=dataset["dev"],
)
trainer.train()

# 儲存訓練完的模型
trainer.save_model()