mergekit saves tied and ignored weights unlike what transformers does when saving

[nyxkrage]

An example of this in the wild is Gemma2, where saving a Gemma2 model with save_pretrained ignores the lm_head tensor, due to it being a tied weight, whereas a Gemma2 model saved from mergekit will have the lm_head included in the safetensors files, resulting in the model being 10.2B parameters vs 9.24B parameters. This can be seen with grimjim/Gemma2-Nephilim-v3-9B vs google/gemma-2-9b

Relevant code from transformers:
https://github.com/huggingface/transformers/blob/3d8bd11942cec26851c80c01aa5e8403542ca50b/src/transformers/modeling_utils.py#L2634-L2637
https://github.com/huggingface/transformers/blob/3d8bd11942cec26851c80c01aa5e8403542ca50b/src/transformers/modeling_utils.py#L2666-L2700