Some group norm input varation failed

[swimdi]

These input varations of group_norm are failed

"input_shape, num_groups",
# These two cases appeared in stable diffusion v2 and accuracy failed
pytest.param((1, 1280, 8, 8), 32, marks=pytest.mark.xfail(reason="")),
pytest.param((1, 2560, 8, 8), 32, marks=pytest.mark.xfail(reason="")),
# These four cases appeared in retinanet_resnet50_fpn_v2 and RuntimeError
pytest.param((1, 256, 50, 68), 32, marks=pytest.mark.xfail(reason="")),
pytest.param((1, 256, 25, 34), 32, marks=pytest.mark.xfail(reason="")),
pytest.param((1, 256, 13, 17), 32, marks=pytest.mark.xfail(reason="")),
pytest.param((1, 256, 7, 9), 32, marks=pytest.mark.xfail(reason="")),

You can reproduce by

1. remove aten_native_group_norm_default_blocklist
2. pytest tests/lowering/normalization/test_group_norm.py --runxfail

Btw, if the input_tensor layout set TtnnTileLayout then this two case accuracy failed

"input_shape, num_groups",
[(1, 320, 32, 32), 32]
[(2, 320, 64, 64), 32]