[Bugfix] Fix AutoModel.from_pretrained(..., quantization_config=None) regression

[kylesayrs]

Purpose

• Fix bug where passing quantization_config=None to AutoModelForCausalLM will lead to an error

Traceback (most recent call last):
  File "/home/kyle/llm-compressor/asdf.py", line 2, in <module>
    model = AutoModelForCausalLM.from_pretrained("RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w4a16", quantization_config=None)
  File "/home/kyle/transformers/src/transformers/models/auto/auto_factory.py", line 547, in from_pretrained
    config, kwargs = AutoConfig.from_pretrained(
  File "/home/kyle/transformers/src/transformers/models/auto/configuration_auto.py", line 1277, in from_pretrained
    return config_class.from_dict(config_dict, **unused_kwargs)
  File "/home/kyle/transformers/src/transformers/configuration_utils.py", line 817, in from_dict
    logger.info(f"Model config {config}")
  File "/home/kyle/transformers/src/transformers/configuration_utils.py", line 851, in __repr__
    return f"{self.__class__.__name__} {self.to_json_string()}"
  File "/home/kyle/transformers/src/transformers/configuration_utils.py", line 963, in to_json_string
    config_dict = self.to_diff_dict()
  File "/home/kyle/transformers/src/transformers/configuration_utils.py", line 865, in to_diff_dict
    config_dict = self.to_dict()
  File "/home/kyle/transformers/src/transformers/configuration_utils.py", line 942, in to_dict
    self.quantization_config.to_dict()
AttributeError: 'NoneType' object has no attribute 'to_dict'

• This option was supported in previous releases, and is documented in docstring

quantization_config (`Union[QuantizationConfigMixin,Dict]`, *optional*):

• I'm not exactly sure what change caused the regression, still trying to track that down

Changes

• Ensuring that the quantization config is never overwritten when instantiating from _BaseAutoModelClass

Testing

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w4a16", quantization_config=None)

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: auto

[Rocketknight1]

cc @MekkCyber for quantization maybe? Feel free to ping me if you think it should go to someone else