Eg: on kaggle where for some reason it holds onto high RAM so model merges arent possible as it runs out of memory.