Unable to merge lora adapter with base model, after doing CPT on a model with tied_embeddings=True

[winter cave]

I did SFT , 16 bit LoRA on qwen 3 vl 2b instruct. The adapters are saved, but on trying to merge,

model, tokenizer = FastVisionModel.from_pretrained("/workspace/qwen2b-lora-cpt-75p-equal-adapters/checkpoint-9500",load_in_4bit=True)
==((====))== Unsloth 2026.2.1: Fast Qwen3_Vl patching. Transformers: 4.57.6.
\ /| NVIDIA RTX A6000. Num GPUs = 1. Max memory: 44.449 GB. Platform: Linux.
O^O/ _/ \ Torch: 2.10.0+cu128. CUDA: 8.6. CUDA Toolkit: 12.8. Triton: 3.6.0
\ / Bfloat16 = TRUE. FA [Xformers = 0.0.35. FA2 = False]
"-____-" Free license: http://github.com/unslothai/unsloth
Unsloth: Fast downloading is enabled - ignore downloading bars which are red colored!

/venv/train/lib/python3.11/site-packages/peft/tuners/tuners_utils.py:919: UserWarning: Model with tie_word_embeddings=True and the tied_target_modules=['model.language_model.embed_tokens', 'lm_head'] are part of the adapter. This can lead to complications, for example when merging the adapter or converting your model to formats other than safetensors. See for example https://github.com/huggingface/peft/issues/2018.

model.save_pretrained_merged(save_directory="qwen-cpt-merged",tokenizer = tokenizer)

Trucating previous stack trace,
...
File "/venv/train/lib/python3.11/site-packages/unsloth_zoo/saving_utils.py", line 1418, in get_torch_storage_size_new
shape = (x.module.in_features, x.module.out_features)
^^^^^^^^^^^^^^^^^^^^
File "/venv/train/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1965, in getattr
raise AttributeError(
AttributeError: 'Embedding' object has no attribute 'in_features'

firstly i do not understand, if the lm head and embeddings are tied (i.e transposes of each other) and we attatch adapter either 1 for both or 1 each, why should the merge have issues? I did not add any new tokens or resize the embedddings

[winter cave]

Any help guys?