release GPU memory

[zhaohai1314]

i made a project which has VUE front end, and the functions are as follows:

1. train codes part : responsible for checking the VUE if there is any training task, if true,then give a thread to train ;

and : a). my question is , the training process may be interrupted by VUE front end, and the GPU memory is occupied about 2+G,which can not be released, i have use the :torch.cuda.empty_cache. what can i do to totally release the remain momory?
b). and if the training process is interrupted once, if second task is coming, it will occur below error