Update app.py

app.py

@@ -16,7 +16,7 @@ def load_models():
     
     model = AutoModelForCausalLM.from_pretrained(
         model_name,
-        torch_dtype=torch_bfloat16,
+        torch_dtype=torch.bfloat16,
         # device_map="auto"
     )
     tokenizer = AutoTokenizer.from_pretrained(model_name)