Fix: propagate quantization_config to text sub-config for composite models in AutoModelForCausalLM

[lvliang-intel]

What does this PR do?

Fixes loading of quantized composite models (e.g. Qwen3.5-35B-A3B with AutoRound quantization) via AutoModelForCausalLM.from_pretrained.

Problem:
For composite models whose model_class.config_class maps to text_config, the from_pretrained method in auto_factory.py swaps the full composite config with the text sub-config:

if model_class.config_class == config.sub_configs.get("text_config", None):
    config = config.get_text_config()

The quantization_config is stored on the top-level composite config (from config.json), but not on the text sub-config. When modeling_utils.py later calls get_hf_quantizer, it checks hasattr(config, "quantization_config") to determine pre_quantized. Since the text sub-config lacks this attribute, pre_quantized is set to False, which causes a ValueError for quantization methods that require pre-quantized weights (e.g. AutoRound, GPTQ, AWQ).

ValueError: The quantization method QuantizationMethod.AUTOROUND does require the model to be pre-quantized.
You explicitly passed pre_quantized=False meaning your model weights are not quantized.

How to reproduce the issue:

qconfig = transformers.AutoRoundConfig()
model = transformers.AutoModelForCausalLM.from_pretrained(model_name_or_path="Intel/Qwen3.5-35B-A3B-int4-AutoRound", trust_remote_code=True, device_map="cuda", quantization_config=qconfig)

Fixes # (issue)

After swapping to the text sub-config, propagate quantization_config from the parent composite config if it exists and the text sub-config does not already have one.

[Rocketknight1]

cc @SunMarc for quants

[SunMarc]

Thanks a lot ! Left a suggestion and a comment

[github-actions]

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

run-slow: auto

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[SunMarc]

Nice !