Update app.py

app.py

@@ -128,7 +128,7 @@ def infer(genre_txt_content, lyrics_txt_content, num_segments, max_new_tokens):
         "--output_dir", f"{output_dir}",
         "--cuda_idx", "0",
         "--max_new_tokens", f"{max_new_tokens}",
-        #"--disable_offload_model"
+        "--disable_offload_model"
     ]
 
     # Set up environment variables for CUDA with optimized settings