Update app.py

app.py

@@ -16,66 +16,64 @@ tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
 zero = torch.Tensor([0]).cuda()
 print(zero.device) # <-- 'cpu' 🤔
 
-def greet(n):
-    print(zero.device) # <-- 'cuda:0' 🤗
-        
-    
-    print(zero.device) # <-- 'cuda:0' 🤗
-    
-    
-    
-    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)
+
+print(zero.device) # <-- 'cuda:0' 🤗
     
-    # === 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
+
+print(zero.device) # <-- 'cuda:0' 🤗
+
+
+
+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 ===
+@spaces.GPU
+def trainf():
+
+    v= Trainer(
     
-    # === 5️⃣ TOKENLEŞTİRME FONKSİYONU ===
-    def tokenize_function(examples):
-        return tokenizer(examples["text"], truncation=True, max_length=512)
+        model=model,
     
-    tokenized_datasets = train_data.map(tokenize_function, batched=True)
+        args=training_args,
     
-    # === 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
+        train_dataset=tokenized_datasets,
+
     )
-    
-    # === 7️⃣ MODEL EĞİTİMİ ===
-    @spaces.GPU
-    def trainf():
-    
-        v= Trainer(
-        
-            model=model,
-        
-            args=training_args,
-        
-            train_dataset=tokenized_datasets,
-   
-        )
-        return v
-    
-    trainf().train()
-demo = gr.Interface(fn=greet)
-demo.launch()
+    return v
+
+trainf().train()