add_batch taking very long time (SegFormer semantic segmentation).

When I call **"add_batch"** with 512X512 prediction image (and 512X512) label
With batch size of 40 (so the input are 40X512X512).
It takes **50 seconds per batch** on my machine (windows 10 with RTX 3090 and i9-12900K intel CPU. Torch: 1.12.1+cu113).
(It is stuck in "encode_batch" which call the recursive "encode_nested_example")
It is much longer than training itslef

As a patch:
**I see that If I switch to "Evaluate.compute" it still takes long time but "Evaluate._compute" is much faster.**

**Can I change "add_batch" to something else to enjoy the faster execution the same way replacing "compute" with "_compute ?**