Update app.py

app.py

@@ -9,7 +9,6 @@ if not huggingface_token:
     raise ValueError("HUGGINGFACE_TOKEN environment variable is not set")
 
 model_id = "meta-llama/Llama-Guard-3-8B-INT8"
-device = "cuda" if torch.cuda.is_available() else "cpu"
 dtype = torch.bfloat16
 
 quantization_config = BitsAndBytesConfig(load_in_8bit=True)
@@ -38,6 +37,7 @@ def parse_llama_guard_output(result):
 
 @spaces.GPU
 def moderate(user_input, assistant_response):
+    device = "cuda" if torch.cuda.is_available() else "cpu"
     tokenizer = AutoTokenizer.from_pretrained(model_id, token=huggingface_token)
     model = AutoModelForCausalLM.from_pretrained(
         model_id,