Will the self-compiled operator be more stable when infering

[brilliant-soilder]

Will the self-compiled operator be more stable than the non-self-compiled cuda tensorrt in infering, will the infering be affected by the automatic upgrade of nvidia driver, or cuda version, so as to become unstable? And will the specific operator determines the apecific GPU resources used.