`.to_dict` does not correctly serialize `torch.dtype` in some cases (e.g., vision models)

[xenova]

System Info

• transformers version: 4.29.1
• Platform: Windows-10
• Python version: 3.8.3
• Huggingface_hub version: 0.14.1
• Safetensors version: 0.3.1
• PyTorch version (GPU?): 2.0.0+cu117 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: no
• Using distributed or parallel set-up in script?: no

Who can help?

@sgugger @ArthurZucker @amyeroberts

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

Running:

import json
from transformers import AutoConfig
json_data = AutoConfig.from_pretrained('openai/clip-vit-base-patch16').to_dict()
json.dumps(json_data, indent=4)

Results in

TypeError: Object of type dtype is not JSON serializable

---

I have identified this problem with the following models:

• clip
• sam
• vision-encoder-decoder

Expected behavior

torch dtypes should be converted to a string. I believe this is due to these configs redefining their to_dict method, without calling dict_torch_dtype_to_str on the top-level object.

transformers/src/transformers/models/clip/configuration_clip.py

Lines 397 to 408 in de9255d

	def to_dict(self):
	"""
	Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
	Returns:
	`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
	"""
	output = copy.deepcopy(self.__dict__)
	output["text_config"] = self.text_config.to_dict()
	output["vision_config"] = self.vision_config.to_dict()
	output["model_type"] = self.__class__.model_type
	return output

[ArthurZucker]

Hey! Indeed, the PretrainedConfig class calls dict_torch_dtype_to_str, and the text_config and vision_config inherit from it, so they work fine, indeed, the parent's torch_dtype attribute can be modified and we don't use self.to_dict() . Thanks for reporting

The configs should be automatically tested IMO, this is currently note the case. It seems that for blip, only the text config is tested, which is why this does not fail. 10 models or more are concerned (mostly when is_composition=True.

I'll open a PR to fix this

[xenova]

Commenting to prevent it being closed as stale.

[ArthurZucker]

Yep, sorry I'll try to get to the original fix taking the comment into account!

[ArthurZucker]

This was fixed by #25237