Update app.py

app.py

@@ -70,9 +70,9 @@ def respond(
         if llm is None or llm_model != model:
             llm = Llama(
                 model_path=f"models/{model}",
-                flash_attn=True,
-                n_gpu_layers=81,
-                n_batch=1024,
+                flash_attn=False,
+                n_gpu_layers=0,
+                n_batch=32,
                 n_ctx=8192,
             )
             llm_model = model