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
import time
import spaces

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

torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch_dtype, device_map="auto")

# === 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).to("cuda" if torch.cuda.is_available() else "cpu")

# === 3️⃣ VERİ SETİ ===
@spaces.GPU
def load_and_prepare_dataset():
    dataset = load_dataset("oscar", "unshuffled_deduplicated_tr", trust_remote_code=True)
    subset = dataset["train"].shuffle(seed=42).select(range(10000))
    
    def tokenize_function(examples):
        return tokenizer(examples["text"], truncation=True, max_length=512)
    
    tokenized_datasets = subset.map(tokenize_function, batched=True)
    return tokenized_datasets

tokenized_dataset = load_and_prepare_dataset()

# === 4️⃣ EĞİTİM AYARLARI ===
batch_size = 1  
num_epochs = 1  
max_steps = (len(tokenized_dataset) // batch_size) * num_epochs

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,  
    save_steps=500,
    save_total_limit=2,
    logging_dir="./logs",
    logging_steps=10,
    optim="adamw_torch",
    fp16=torch.cuda.is_available(),  
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
)

@spaces.GPU
def train_model():
    trainer.train()

# === 5️⃣ CHAT ARAYÜZÜ ===
@spaces.GPU
def slow_echo(message, history):
    response = "Model henüz eğitilmedi. Lütfen eğitimi başlatın."
    if model:
        response = f"You typed: {message}"
    return response

demo = gr.ChatInterface(
    slow_echo,
    type="text",
    flagging_mode="manual",
    flagging_options=["Like", "Spam", "Inappropriate", "Other"],
    save_history=True,
)

if __name__ == "__main__":
    train_model()
    demo.launch(share=True)