import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from urllib.parse import unquote_plus
import os

from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed, Trainer, TrainingArguments, BitsAndBytesConfig, \
    DataCollatorForLanguageModeling, Trainer, TrainingArguments
from transformers import BitsAndBytesConfig
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training

nf4_config = BitsAndBytesConfig(
   load_in_4bit=True,
   bnb_4bit_quant_type="nf4",
   bnb_4bit_use_double_quant=True,
   bnb_4bit_compute_dtype=torch.bfloat16
)

# Carregar o modelo e o tokenizador na GPU
device = "cuda:0"
model_id = "meta-llama/Meta-Llama-3-8B-Instruct"
model = AutoModelForCausalLM.from_pretrained(model_id,quantization_config=nf4_config,device_map="auto",local_files_only=False,trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_id, use_default_system_prompt=False)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
print(model)
from transformers import AutoModelForCausalLM
from datasets import load_dataset
from trl import *

# jondurbin/truthy-dpo-v0.1

def return_prompt_and_responses(samples) :
    return {
        "prompt": [
            "Question: " + question + "\n\nAnswer: "
            for question in samples["prompt"]
        ],
        "chosen": samples["chosen"],   # rated better than k
        "rejected": samples["rejected"], # rated worse than j
    }

dataset = load_dataset(
    "jondurbin/truthy-dpo-v0.1",
    split="train",
    #data_dir="data/rl"
)
original_columns = dataset.column_names

dataset.map(
    return_prompt_and_responses,
    batched=True,
    remove_columns=original_columns
)


model = prepare_model_for_kbit_training(model)

peft_config = LoraConfig(
 r=128,
 lora_alpha=16,
 target_modules=["q_proj","k_proj","v_proj","o_proj", "up_proj","gate_proj","down_proj", "lm_head"],
 lora_dropout=0.05,
 bias="none",
 task_type="CAUSAL_LM",
)
output_dir = "./odp"
training_args = TrainingArguments(
 per_device_train_batch_size=1,
 gradient_accumulation_steps=1,
 gradient_checkpointing =True,
 max_grad_norm= 0.3,
 optim='adafactor',
 overwrite_output_dir=True,save_steps=100,
 num_train_epochs=1,
 learning_rate=2e-4,
 bf16=True,
 save_total_limit=3,
 logging_steps=10,
 output_dir=output_dir,
 lr_scheduler_type="cosine",
 warmup_ratio=0.05,
)

dpo_trainer = DPOTrainer(
 model,
 #model_ref,
 args=training_args,
 peft_config=peft_config,
 beta=0.1,
 train_dataset=dataset,
 tokenizer=tokenizer,
 max_prompt_length=1024,
 max_length=2048,
)

dpo_trainer.train()