I ran into an out of memory error while training a reinforcement learning network.

The error says that 11 Gb is already allocated and 13Gb is reserved by pytorch
However when running nvidia-smi I see only 3Gb being used

Is there a way to know from where the additional allocated memory comes from ?