Function swap_scale_shift

diffusers/src/diffusers/loaders/single_file_utils.py:1667–1670

(weight, dim)

# in SD3 original implementation of AdaLayerNormContinuous, it split linear projection output into shift, scale;
# while in diffusers it split into scale, shift. Here we swap the linear projection weights in order to be able to use diffusers implementation
def swap_scale_shift(weight, dim):
    shift, scale = weight.chunk(2, dim=0)
    new_weight = torch.cat([scale, shift], dim=0)
    return new_weight


def convert_sd3_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):