Update app.py

app.py

@@ -40,7 +40,7 @@ def respond(
     if llm is None or llm_model != model:
         llm = Llama(
             model_path=f"models/{model}",
-            n_gpu_layers=50,  
+            n_gpu_layers=0,  
             n_batch=32000,     
             n_ctx=2048,      
         )