app.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model
from datasets import load_dataset
import gradio as gr

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

# === 2️⃣ CPU OPTİMİZASYONU ===
torch_dtype = torch.float32  # CPU için uygun dtype
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch_dtype)

# === 3️⃣ 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)

# === 4️⃣ VERİ SETİ ===
dataset = load_dataset("oscar", "unshuffled_deduplicated_tr", trust_remote_code=True)  # trust_remote_code=True
subset = dataset["train"].shuffle(seed=42).select(range(10000))  # Küçük subset seçiyoruz (10.000 örnek)

# === 5️⃣ TOKENLEŞTİRME FONKSİYONU ===
@spaces.GPU
def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, max_length=512)

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

# === 6️⃣ EĞİTİM AYARLARI ===
# Eğitimde kaç adım olduğunu hesaplayalım
train_size = len(tokenized_datasets)  # 10,000 örnek
batch_size = 1  # Batch size 1
num_epochs = 1  # 1 epoch eğitimi
max_steps = (train_size // batch_size) * num_epochs  # max_steps hesapla

training_args = TrainingArguments(
    output_dir="./mistral_lora",
    per_device_train_batch_size=1,  
    gradient_accumulation_steps=16, 
    learning_rate=5e-4,  
    num_train_epochs=1,  
    max_steps=max_steps,  # Buraya max_steps parametresini ekliyoruz
    save_steps=500,
    save_total_limit=2,
    logging_dir="./logs",
    logging_steps=10,
    optim="adamw_torch",
    train_dataset=split_dataset["train"],
    eval_dataset=split_dataset["test"],
    no_cuda=True,  # GPU kullanılmıyor
)

# === 7️⃣ MODEL EĞİTİMİ ===
@spaces.GPU
def train_model():
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized_datasets,
    )
    trainer.train()

train_model()  # Eğitimi başlat