Cuda Tensors seems to leak memory in GPU

I notice that the GPU tensors leak memory. When I use Tensors over CPU there is not problem but when I switch to GPU, the memory allocation in GPU begins to accumulate. I am running this on jupyter notebook.

Whenever I interrupt the code while running, the allocated memory on GPU seems to go uncollected.

Is there any guidelines on when to del cuda tensors after usage. I mean the lifecycle of a cuda tensor?