Fix save_pretrained for quantized models with custom serialization

[480284856]

Problem

When calling save_pretrained() on mxfp4 quantized models, a NotImplementedError was raised because revert_weight_conversion() tried to reverse operations that don't implement reverse_op.

Environment

• Transformers Version: Commit a7f29523361b2cc12e51c1f5133d95f122f6f45c (main branch)
• Python Version: 3.10.12
• Model: openai/gpt-oss-20b (mxfp4 quantized)

Reproduction Code

from transformers import AutoModelForCausalLM

# Load the mxfp4 quantized model
model = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b")

# Attempt to save the model
model.save_pretrained("openai/gpt-oss-20b-hf")

Error Traceback

---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
Cell In[1], line 3
      1 from transformers import AutoModelForCausalLM
      2 model = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b")
----> 3 model.save_pretrained("openai/gpt-oss-20b-hf")

File /workspace/transformers/src/transformers/modeling_utils.py:3237, in PreTrainedModel.save_pretrained(self, save_directory, is_main_process, state_dict, push_to_hub, max_shard_size, variant, token, save_peft_format, save_original_format, **kwargs)
   3235 # Revert all renaming and/or weight operations
   3236 if save_original_format:
-> 3237     state_dict = revert_weight_conversion(model_to_save, state_dict)

File /workspace/transformers/src/transformers/core_model_loading.py:1164, in revert_weight_conversion(model, state_dict)
   1161     return state_dict
   1163 # Reverse all Transform to correctly match keys
-> 1164 reverse_weight_conversion = [conversion.reverse_transform() for conversion in weight_conversions]

File /workspace/transformers/src/transformers/core_model_loading.py:1164, in <listcomp>(.0)
-> 1164 reverse_weight_conversion = [conversion.reverse_transform() for conversion in weight_conversions]

File /workspace/transformers/src/transformers/core_model_loading.py:531, in WeightTransform.reverse_transform(self)
   528 # Add the reverse ops if applicable (it needs to be provided at __init__)
   529 if hasattr(self, "operations"):
   530     # All reverse ops, in reverse order
-> 531     kwargs["operations"] = [op.reverse_op for op in self.operations[::-1]]

File /workspace/transformers/src/transformers/core_model_loading.py:531, in <listcomp>(.0)
-> 531     kwargs["operations"] = [op.reverse_op for op in self.operations[::-1]]

File /workspace/transformers/src/transformers/core_model_loading.py:102, in ConversionOps.reverse_op(self)
   100 @property
   101 def reverse_op(self) -> ConversionOps:
-> 102     raise NotImplementedError

NotImplementedError: 

Solution

Skip revert_weight_conversion() when the quantizer has already provided the state_dict via get_state_dict_and_metadata(), since quantizers handle their own serialization logic.

Changes

• Added quantizer_provided_state_dict flag to track when quantizer provides state_dict
• Skip revert_weight_conversion() when quantizer already provided serialized state_dict

Fixes the issue where mxfp4 quantized models cannot be saved due to missing reverse_op implementation.

by Claude Opus 4.5

[ArthurZucker]

cc @SunMarc I think we plan to bring back serialization

[SunMarc]

hmmm indeed, we need to fix this cc @MekkCyber

[MekkCyber]

Hey @480284856 ! thanks for the contribution but i'm not really sure if we want to handle it this way, I think it makes more sense to have a reverse op for quantization too instead of handling it differently using a quantization state dict