Update app.py

app.py

@@ -58,7 +58,6 @@ async def load_models():
     model = AutoModelForCausalLM.from_pretrained(
         MODEL_REPO,
         low_cpu_mem_usage=True,
-        return_legacy_cache=True,
         device_map={"": 0} if device == "cuda" else None,
         torch_dtype=torch.bfloat16 if device == "cuda" else None,
     )