benchmark

5140f9f4 - skip timm_vit_large on cuda CI since it ooms