Update app.py

app.py

@@ -42,7 +42,7 @@ def respond(
     
     if llm is None or llm_model != model:
         llm = Llama(
-            model_path=f"models/{model}",
+            model_path="models/open_gpt4_8x7b_v0.2.Q5_K_M.gguf",
             flash_attn=True,
             n_gpu_layers=81,
             n_batch=1024,