System Info

• transformers version: 4.57.0.dev0
• Platform: Linux-5.15.0-153-generic-x86_64-with-glibc2.31
• Python version: 3.12.9
• Huggingface_hub version: 0.34.4
• Safetensors version: 0.6.2
• Accelerate version: 1.4.0
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (accelerator?): 2.8.0+cu128 (NA)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?: no

Who can help?

No response

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

After investigation of an issue in trl, I found a weird behavior of transformers config rope_scaling: the config.rope_scaling (at the root config level) and config.text_config.rope_scaling (under text_config) might be the same or different dict objects depending on whether we pass text_config param to AutoConfig.from_pretrained

• if we don't pass text_config param, the 2 rope_scaling point to the same dict object
• if we pass text_config param, the 2 rope_scaling are different dict objects

In [1]: from transformers import AutoConfig

In [2]: model_id = "Qwen/Qwen2.5-VL-3B-Instruct"

In [3]: config1 = AutoConfig.from_pretrained(model_id)
In [4]: config1.text_config.rope_scaling
Out[4]: {'type': 'default', 'mrope_section': [16, 24, 24], 'rope_type': 'default'}
In [5]: config1.rope_scaling
Out[5]: {'type': 'default', 'mrope_section': [16, 24, 24], 'rope_type': 'default'}
In [6]: id(config1.text_config.rope_scaling)
Out[6]: 140211029392000
In [7]: id(config1.rope_scaling)
Out[7]: 140211029392000
# Both are the same dict object

In [8]: config2 = AutoConfig.from_pretrained(model_id, text_config={})
In [9]:config2.text_config.rope_scaling
Out[9]: {'type': 'default', 'mrope_section': [16, 24, 24], 'rope_type': 'default'}
In [10]: config2.rope_scaling
Out[10]: {'type': 'default', 'mrope_section': [16, 24, 24], 'rope_type': 'default'}
In [11]: id(config2.text_config.rope_scaling)
Out[11]: 140210801100608
In [12]: id(config2.rope_scaling)
Out[12]: 140211029786688
# Both are different dict objects

Is this expected?

We discovered this while investigating why changing (after initialization) the config.text_config.rope_scaling will or will not change the config.rope_scaling as well. See related comment in trl PR:

• Fix VLM configs in generate_tiny_models trl#4101 (comment)

Expected behavior

• Either they should be the same dict object in any case
• Or they should be different dict objects in any case