diff --git a/mz_llama3.py b/mz_llama3.py
index e842e09..2b3d7ef 100644
--- a/mz_llama3.py
+++ b/mz_llama3.py
@@ -146,5 +146,6 @@ def query_beautify_prompt_text(model_name, n_gpu_layers, text, style_presets, do
         return full_response
 
     except Exception as e:
+        mz_llama_cpp.freed_gpu_memory(model_file=model_file)
         # mz_utils.Utils.print_log(f"Error in auto_prompt_text: {e}")
         raise e
\ No newline at end of file