Update app.py

app.py

@@ -22,8 +22,6 @@ token = os.environ["HF_TOKEN"]
 
 model = AutoModelForCausalLM.from_pretrained(
     "microsoft/Phi-3-mini-128k-instruct",
-    use_cache=False,
-    attn_implementation="flash_attention_2",
     token=token,
     trust_remote_code=True,
 )