Question about tie_embeddings

I have a couple of questions about tie_embeddings but I don't have enough experience with lightning. I'm sorry if I'm mistaken.

1. In litgpt/pretrain.py, model.transformer.wte.weight = model.lm_head.weight is applied when tie_embeddings is true.
   In my observation, after training, lm_head.weight have tuned parameters while transformer.wte.weight is kept unchanged in the checkpoint file.
   However, in litgpt/scripts/convert_lit_checkpoint.py, transformer.wte.weight is saved while lm_head.weight is skipped.
   So, I'm suspecting conversion to huggingface format is incorrect.

2. How to enable tie_embedding for finetune_full? Using the same way as pretrain is not effective?

3. Is it better to move tie_embedding from train config to model config? There are many chances to mistake setting of tie_embeddings such as finetuning non-tied pretrained as tied, and vice versa.

Thank you for everything.

Start by reading litgpt/pretrain.py and litgpt/scripts/convert_lit_checkpoint.py to trace how tied weights are assigned and which checkpoint key is converted. Then inspect the finetune_full entry point and related train/model configuration handling. Done means establishing whether conversion preserves tied embeddings and documenting or implementing consistent behavior for pretraining and finetuning.