[Regression] Qwen3.5 saved checkpoint is not correct with `save_pretrained` API since version 5.4.0

[xin3he]

System Info

transformers == 5.3.0 works well
transformers ==5.4.0 returns Unexpected model.language_model.language_model.language_model.layers.7.self_attn.v_proj.weight in loaded safetensors file

Who can help?

@zucchini-nlp

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

import transformers

name = "Qwen/Qwen3.5-0.8B"
model = transformers.Qwen3_5ForConditionalGeneration.from_pretrained(name, trust_remote_code=True)
model.save_pretrained("./qwen-35")

from safetensors.torch import save_file, load_file

loaded = load_file("./qwen-35/model.safetensors")
assert not 'model.language_model.language_model.language_model.layers.7.self_attn.v_proj.weight' in loaded, "Unexpected model.language_model.language_model.language_model.layers.7.self_attn.v_proj.weight in loaded safetensors file"

Expected behavior

assert pass

[zucchini-nlp]

Oh no, some conversions are meant to be applied only for certain classes. A quick workaround could be to bring back the cls attribute for conversion_mapping though as I mentioned earlier to @Cyrilvallez imo we need a better way to store conversions. We have several regex patterns that are model-class specific, for ex applied only when a prediction is present

Ig in long term we have to adjust conversion dict's layout slightly