Inference GPU memory is more than training

[ker95qq]

I used internimage base segmentation model, however GPU memory of inference is more than training (batch size 1). I am very confused.