Suddenly CUDA out of memory

[6gujn]

When I combined robust_loss with the Informer model and set all parameters to a minimum, the model memory suddenly increased at runtime (as shown in the image above). And running the combined model on a 16GB graphics card is the same problem

[6gujn]

With the same parameter Settings, this problem would not have occurred if only the Informer model was used (above is the memory result with only the Informer model). Does anyone have this problem? How was it resolved?

[6gujn]

This issue has been resolved