Update app.py

app.py

@@ -933,12 +933,13 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))</pre>
 # ==================== MODEL YÜKLEME ====================
 print("Modeller yükleniyor... Bu biraz sürebilir (özellikle ilk seferde).")
 
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
 model_stable = AutoModelForCausalLM.from_pretrained(
     "sixfingerdev/kayra-1",
     trust_remote_code=True,
-    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
-    device_map="auto"
-)
+    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
+).to(device)
 tokenizer_stable = AutoTokenizer.from_pretrained("sixfingerdev/kayra-1")
 
 model_exp = AutoModelForCausalLM.from_pretrained(