NVIDIAModelOptConfig.to_dict() crashes from_pretrained/save_pretrained when forward_loop is set #15122

Describe the bug

forward_loop is a documented NVIDIAModelOptConfig constructor parameter. NVIDIAModelOptConfig has no to_dict() override, so it inherits QuantizationConfigMixin.to_dict(), which returns copy.deepcopy(self.__dict__) — including the raw forward_loop callable.

src/diffusers/models/modeling_utils.py calls json.dumps(hf_quantizer.quantization_config.to_dict(), sort_keys=True) for telemetry during from_pretrained, with no error handling around it. Any model loaded (or saved) with a NVIDIAModelOptConfig(forward_loop=...) crashes with TypeError: Object of type function is not JSON serializable.

Reproduction

import json
from diffusers.quantizers.quantization_config import NVIDIAModelOptConfig

config = NVIDIAModelOptConfig(
    quant_type="FP8",
    forward_loop=lambda model: model,
    modelopt_config={"quant_cfg": {}},
)

json.dumps(config.to_dict())

Logs

Traceback (most recent call last):
  File "<string>", line 1, in <module>
TypeError: Object of type function is not JSON serializable

This is the same call modeling_utils.py makes internally during from_pretrained:

  File ".../diffusers/src/diffusers/models/modeling_utils.py", line 1206, in from_pretrained
    serialized_quant_config = json.dumps(hf_quantizer.quantization_config.to_dict(), sort_keys=True)
TypeError: Object of type function is not JSON serializable

System Info

• diffusers @ main, commit aaf57d0
• Python 3.12, Linux

Who can help?

(leaving blank — no specific quantization reviewer listed in the template's tag guide)