Qwen3.5 lora merge export using wrong keys

[perone]

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

There is a mismatch of the keys when merging the lora adapters for Qwen3.5 27B:

  ┌───────────────────────────────────────────────────────────────┬─────────────────────────────────┐
  │                       Merged checkpoint                       │      Base model (expected)      │
  ├───────────────────────────────────────────────────────────────┼─────────────────────────────────┤
  │ model.language_model.language_model.language_model.layers.X.* │ model.language_model.layers.X.* │
  ├───────────────────────────────────────────────────────────────┼─────────────────────────────────┤
  │ model.language_model.visual.*                                 │ model.visual.*                  │
  ├───────────────────────────────────────────────────────────────┼─────────────────────────────────┤
  │ lm_head.weight                                                │ lm_head.weight (correct)        │
  └───────────────────────────────────────────────────────────────┴─────────────────────────────────┘

That causes issues with vLLM 0.19.0 when loading the exported model. I'm using transformers==5.5.0 and ms-swift commit 0f5fb0083adc7d674c6126bea1a0d2b1e20a1795.

How to Reproduce / 如何复现

Model exported with:

swift export --adapters checkpoint-xxx --merge_lora true --use_hf true

Additional Information / 补充信息

No response