app.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model
from datasets import load_dataset
import spaces
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 ===



zero = torch.Tensor([0]).cuda()
print(zero.device) # <-- 'cpu' 🤔
@spaces.GPU
def greet(n):
    print(zero.device) # <-- 'cuda:0' 🤗
    return f"Hello {zero + n} Tensor"

print(zero.device) # <-- 'cuda:0' 🤗

device = "cpu"  # CPU kullanıyoruz
torch_dtype = torch.float32  # float32 seçtik çünkü CPU'da bf16 genelde yok
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)  # 🔥 ÇÖZÜM: trust_remote_code=True
train_data = dataset["train"].shuffle(seed=42).select(range(10000))  # Küçük subset

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

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

# === 6️⃣ 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",  # 🔥 ÇÖZÜM: bitsandbytes yerine adamw_torch
)

# === 7️⃣ MODEL EĞİTİMİ ===
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets,
)

trainer.train()