memory release

[shimen]

I have a huge difference in the available GPU memory.

• When using the command: nvidia smi
  | 2 Tesla V100-SXM2... On | 00000000:0A:00.0 Off | 0 |
  | N/A 43C P0 70W / 300W | 16101MiB / 16160MiB | 0% Default |

• When using the cutorch command: cutorch.getMemoryUsage(cutorch.getDevice())
  Epoch: [1][2/5] Time 0.724 tripErr 7.65e-02
  -------- exit epoch
  7911792640 16945512448
  -------- start epoch
  7915701248 16945512448
  (nTrips, nTripsFound) = (5206, 369)
  Epoch: [1][3/5] Time 0.505 tripErr 8.91e-02
  -------- exit epoch
  7923902976 16945512448
  -------- start epoch
  7927670784 16945512448
  (nTrips, nTripsFound) = (7141, 265)
  Epoch: [1][4/5] Time 0.623 tripErr 7.40e-02
  -------- exit epoch
  7867307008 16945512448
  -------- start epoch
  7918349824 16945512448
  (nTrips, nTripsFound) = (5658, 378)
  Epoch: [1][5/5] Time 0.469 tripErr 7.18e-02
  -------- exit epoch
  7922003968 16945512448

How come there is such a huge difference?
Almost 0MB free with "nvidia smi" command while the "cutorch" command showing that there is almost 8000mb free.