made changes in max_cache_length

app.py

@@ -50,7 +50,7 @@ demo = gr.Interface(
 # Patch Phi-3.5-mini-instruct's `prepare_inputs_for_generation`
 def patched_prepare_inputs_for_generation(self, input_ids, past_key_values=None, **kwargs):
     if past_key_values is not None:
-        max_cache_length = past_key_values.get_seq_length()  # Fix here
+        max_cache_length = len(past_key_values[0][0])
     else:
         max_cache_length = None
     return self._default_prepare_inputs_for_generation(input_ids, past_key_values, **kwargs)