app.py

import os
import torch
import gradio as gr
import spaces
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model
from datasets import load_dataset, Dataset
from huggingface_hub import notebook_login, HfApi

# === 1️⃣ MODEL VE TOKENIZER YÜKLEME ===
MODEL_NAME = "mistralai/Mistral-7B-v0.1"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch.float32).to(device)

# === 2️⃣ LoRA AYARLARI ===
lora_config = LoraConfig(
    r=8,
    lora_alpha=32,
    lora_dropout=0.1,
    bias="none",
    target_modules=["q_proj", "v_proj"],
)
model = get_peft_model(model, lora_config)

# === 3️⃣ VERİ SETİ ===
DATASET_PATH = "/home/user/app/oscar_tr.parquet"

if os.path.exists(DATASET_PATH):
    print("📂 Kaydedilmiş dataset bulundu, yükleniyor...")
    dataset = Dataset.from_parquet(DATASET_PATH)
else:
    print("🌍 Veri seti indiriliyor ve kaydediliyor...")
    dataset = load_dataset("oscar", "unshuffled_deduplicated_tr", split="train", streaming=True, trust_remote_code=True)
    dataset = dataset.shuffle(seed=42).select(range(10000))  # Küçük subset
    dataset.to_parquet(DATASET_PATH)  # İlk çalışmada kaydediyoruz

# === 4️⃣ TOKENLEŞTİRME ===
def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, max_length=512)

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

# === 5️⃣ EĞİTİM AYARLARI ===
training_args = TrainingArguments(
    output_dir="./mistral_lora_cpu",
    per_device_train_batch_size=1,
    gradient_accumulation_steps=16,
    learning_rate=5e-4,
    num_train_epochs=1,
    save_steps=500,
    save_total_limit=2,
    logging_dir="./logs",
    logging_steps=10,
    optim="adamw_torch",
)

# === 6️⃣ GPU İLE EĞİTİM BAŞLATMA ===
@spaces.GPU
def train_model():
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized_datasets,
    )
    trainer.train()
    return "✅ Model Eğitimi Tamamlandı!"

# === 7️⃣ MODELİ HUGGING FACE HUB'A YÜKLEME ===
def upload_model():
    notebook_login()  # Hugging Face hesabına giriş yap
    api = HfApi()
    api.upload_folder(
        folder_path="./mistral_lora_cpu",
        repo_id="kullanici_adin/mistral-lora-modeli",
        repo_type="model",
    )
    return "✅ Model Hugging Face Hub'a Yüklendi!"

# === 8️⃣ GRADIO ARAYÜZÜ ===
def generate_text(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    output = model.generate(**inputs, max_length=100)
    return tokenizer.decode(output[0], skip_special_tokens=True)

iface = gr.Interface(
    fn=generate_text,
    inputs=gr.Textbox(lines=2, placeholder="Buraya bir şeyler yaz..."),
    outputs="text",
    live=True
)

# === 9️⃣ BAŞLATMA ===
if __name__ == "__main__":
    train_model()  # Eğitimi başlat
    upload_model()  # Modeli Hugging Face Hub'a yükle
    iface.launch()  # Gradio UI başlat