Pytorch cuda memory leak

When I do multiplication matrices and does not assign result any variable it increase gpu memory usage

I can find this tensors in gc objects

but can’t remove its

Same question was here but without answer python - GPU memory usage accumulating when calling function even though tensor isn't stored - Stack Overflow

Why it happen and how find and remove this objects

It is not pytorch, it is IPython. IPython create variables like ___{number of row}. If remove this variable with all references (use %xdel) memory free