import pandas as pd
from datasets import Dataset
from openmind import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer
import torch
from peft import LoraConfig, TaskType, get_peft_model
from transformers import DataCollatorForSeq2Seq
from torch.utils.data import DataLoader
df = pd.read_json('./LLaMA-Factory/data/GLM-4-Instruct-4K.json')
ds = Dataset.from_pandas(df)
print(len(ds))

model_path = 'models'

tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False, trust_remote_code=True)

def process_func(example):
    MAX_LENGTH = 2048    
    input_ids, attention_mask, labels = [], [], []
    instruction = tokenizer(f"<|im_start|>system\n你是一个写作助手，请你根据要求进行文字创作。<|im_end|>\n<|im_start|>user\n{example['instruction'] + example['input']}<|im_end|>\n<|im_start|>assistant\n", add_special_tokens=False)  # 
    response = tokenizer(f"{example['output']}", add_special_tokens=False)
    input_ids = instruction["input_ids"] + response["input_ids"] + [tokenizer.pad_token_id]
    attention_mask = instruction["attention_mask"] + response["attention_mask"] + [1]  # 因为eos token咱们也是要关注的所以 补充为1
    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"] + [tokenizer.pad_token_id]  
    if len(input_ids) > MAX_LENGTH:  # 做一个截断
        input_ids = input_ids[:MAX_LENGTH]
        attention_mask = attention_mask[:MAX_LENGTH]
        labels = labels[:MAX_LENGTH]
    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "labels": labels
    }
tokenized_id = ds.map(process_func, remove_columns=ds.column_names)
## 加载训练模型
model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto",torch_dtype=torch.bfloat16)

config = LoraConfig(
    task_type=TaskType.CAUSAL_LM, 
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj"],
    inference_mode=False, # 训练模式
    r=8, # Lora 秩
    lora_alpha=16, # Lora alaph，具体作用参见 Lora 原理
    lora_dropout=0.1# Dropout 比例
)

model = get_peft_model(model, config)

model.enable_input_require_grads()
lora_path = "./output/YI-1.5-9B"

args = TrainingArguments(
    output_dir=lora_path,
    per_device_train_batch_size=1,
    gradient_accumulation_steps=1,
    logging_steps=1,
    num_train_epochs=2,
    logging_dir = './output/YI-1.5-9B/logs',
    logging_nan_inf_filter=True,
    save_steps=500,
    learning_rate=1e-4,
    save_strategy='steps',
    dataloader_num_workers=0,
    gradient_checkpointing=True
)

data_loader = DataLoader(tokenized_id, batch_size=args.per_device_train_batch_size, num_workers=0)

trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized_id,
    data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True)
)

trainer.train()