Need a simple way to gracefully catch CUDA and or HuggingFace errors if they occur due to things like bad model prompting, OOM etc.