fix: implement Mxfp4Dequantize.reverse_op for save_pretrained support

[Hyungkeun-Park-Nota]

What does this PR do?

Fixes save_pretrained() failure for GPT-OSS models loaded with Mxfp4Config(dequantize=True).

When Triton/kernels are unavailable, transformers automatically falls back to dequantize=True, converting MXFP4 weights to bf16. However, save_pretrained() then fails because Mxfp4Dequantize.reverse_op raises NotImplementedError.

Since dequantized models are regular bf16 models, the correct behavior is to save them as-is rather than re-quantize to MXFP4.

Changes

src/transformers/integrations/mxfp4.py:

• Add Mxfp4IdentityOp as Mxfp4Dequantize.reverse_op — passes through bf16 weights unchanged during save

src/transformers/quantizers/quantizer_mxfp4.py:

• Remove quantization_config from model config in _process_model_after_weight_loading when dequantize=True, so the saved config.json does not contain quant_method: mxfp4. Without this, reloading the saved bf16 model would attempt the MXFP4 loading path and fail because _blocks/_scales keys don't exist.

Reproduction

from transformers import AutoModelForCausalLM, Mxfp4Config

# Triton unavailable → auto dequantize fallback
model = AutoModelForCausalLM.from_pretrained(
    "openai/gpt-oss-20b",
    quantization_config=Mxfp4Config(dequantize=True),
)
model.save_pretrained("/tmp/test")  # NotImplementedError before this fix

transformers/core_model_loading.py:629 in <listcomp>
    kwargs["operations"] = [op.reverse_op for op in self.operations[::-1]]
transformers/core_model_loading.py:101 in reverse_op
    raise NotImplementedError

[Rocketknight1]

cc @SunMarc for quantization maybe?

[SunMarc]

model = AutoModelForCausalLM.from_pretrained(
"openai/gpt-oss-20b",
quantization_config=Mxfp4Config(dequantize=True),
)
model.save_pretrained("/tmp/test") # NotImplementedError before this fix

hmmm, this shouldn't trigger a reverse ops when we dequantized the model. I think the right behavior here would be to just save the model in its dequantized form.

[Hyungkeun-Park-Nota]

@SunMarc Thanks for the review! Updated the PR based on your feedback:

1. Removed re-quantization logic — replaced Mxfp4ReverseDequantize with Mxfp4IdentityOp that simply passes through bf16 weights as-is during save
2. Remove quantization_config after dequantize — in _process_model_after_weight_loading, when dequantize=True, we delete model.config.quantization_config so the saved model loads as a regular bf16 model without triggering the MXFP4 loading path

[github-actions]

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

run-slow: mxfp4