import torch
from datasets import load_dataset
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    TrainingArguments,
)
from peft import LoraConfig, get_peft_model
from trl import SFTTrainer
import os

# ============================================
# 0. Préparation env
# ============================================
os.environ["OMP_NUM_THREADS"] = "8"

# 1. Configurations
base_model = "mistralai/Mistral-7B-Instruct-v0.3"
new_model_dir = "./mistral-7b-brvm-finetuned"
output_dir = "./results"

# 2. Device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Utilisation du périphérique: {device}")
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)} - "
          f"Mémoire: {torch.cuda.get_device_properties(0).total_memory / (1024**3):.2f} GB")

# 3. Dataset
dataset = load_dataset("lamekemal/brvm_finetune")

# 4. Charger modèle + tokenizer (FP16 natif, pas de quantization)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)
model.config.use_cache = False
model.gradient_checkpointing_enable()

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

# ============================================
# 5. Preprocessing (max_seq_length=512)
# ============================================
def tokenize_function(examples):
    texts = [
        f"Instruction: {instr}\nRéponse: {resp}"
        for instr, resp in zip(examples["instruction"], examples["response"])
    ]
    return tokenizer(
        texts,
        truncation=True,
        padding="max_length",
        max_length=512,
    )

tokenized_datasets = dataset.map(tokenize_function, batched=True)

# ============================================
# 6. LoRA config (tu peux augmenter r à 64 sur L40S)
# ============================================
lora_config = LoraConfig(
    r=64,
    lora_alpha=128,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj",
                    "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
)
model = get_peft_model(model, lora_config)

# ============================================
# 7. TrainingArguments
# ============================================
training_args = TrainingArguments(
    output_dir=output_dir,
    num_train_epochs=3,
    per_device_train_batch_size=8,   # plus gros batch, tu peux tester 16
    gradient_accumulation_steps=2,
    optim="adamw_torch_fused",       # optimiseur rapide pour gros GPU
    save_steps=100,
    logging_steps=10,
    learning_rate=2e-5,              # plus bas qu’en 4bit
    fp16=True,
    max_grad_norm=1.0,
    warmup_ratio=0.03,
    lr_scheduler_type="cosine",
    report_to="tensorboard",
    eval_strategy="steps",
    eval_steps=100,
    save_total_limit=2,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
)

# ============================================
# 8. Trainer
# ============================================
trainer = SFTTrainer(
    model=model,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["validation"],
    peft_config=lora_config,
    args=training_args
)

# 9. Fine-tuning
trainer.train()

# 10. Sauvegarde locale
trainer.save_model(new_model_dir)

# 11. Push du modèle + tokenizer sur Hugging Face Hub
trainer.push_to_hub("lamekemal/mistral-7b-brvm-finetuned")
tokenizer.push_to_hub("lamekemal/mistral-7b-brvm-finetuned")

print(f"✅ Modèle LoRA sauvegardé localement dans {new_model_dir} et poussé sur Hugging Face Hub")