Update app.py

app.py

@@ -39,7 +39,7 @@ def build_chain():
     model = AutoModelForCausalLM.from_pretrained(
         model_id,
         device_map="auto",          # menjaš u "auto" ako imaš GPU space
-        torch_dtype=torch.float32
+        torch_dtype=torch.float16
     )
 
     pipe = pipeline(