AutoConfig.from_dict Missing in transformers==4.51.3 — Incompatibility with optimum==1.26.1

[rratnakar09]

System Info

I am running into a critical compatibility issue between optimum and recent versions of transformers.

❗ Error Summary
When using:
transformers==4.51.3
optimum==1.26.1
onnx==1.17.0
onnxruntime==1.20.0

The following runtime error is thrown when attempting to load an ONNX model using ORTModelForTokenClassification.from_pretrained:

AttributeError: type object 'AutoConfig' has no attribute 'from_dict'

This traces back to:
config = AutoConfig.from_pretrained(...)
# ↓ internally calls:
return CONFIG_MAPPING[pattern].from_dict(config_dict, **unused_kwargs)

However, in transformers>=4.48, the method AutoConfig.from_dict appears to have been deprecated or removed. This causes optimum to break at runtime when trying to load ONNX models.

📦 Package Versions
transformers - 4.51.3
optimum - 1.26.1
onnx - 1.17.0
onnxruntime - 1.20.0
torch - 2.2.6

Due to a security advisory, we're required to upgrade to transformers>=4.48. However, even with the latest optimum==1.26.1, it appears optimum is not yet updated for compatibility with changes introduced in recent transformers versions.

ASK:
Is support for transformers>=4.48 (particularly 4.51.3) planned in an upcoming optimum release?
Could this AutoConfig.from_dict dependency be refactored or conditionally patched to restore compatibility?
Is there a compatibility roadmap available between transformers and optimum for ONNX workflows?

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction (minimal, reproducible, runnable)

Use transformers==4.51.3 and optimum==1.26.1

Load an exported ONNX model using ORTModelForTokenClassification.from_pretrained(...)

Observe the AttributeError about AutoConfig.from_dict

Expected behavior

When using optimum==1.26.1 with transformers>=4.48 (specifically 4.51.3), the following should work without error:
from optimum.onnxruntime import ORTModelForTokenClassification
model = ORTModelForTokenClassification.from_pretrained("path/to/onnx/model")

The model should load successfully using the ONNX Runtime backend.

Internally, AutoConfig.from_pretrained(...) should function correctly regardless of changes in the transformers API (e.g., deprecation/removal of from_dict).

ONNX workflows should remain compatible with newer transformers versions, allowing teams to benefit from critical updates and security patches without breaking ONNX integration.

[gidzr]

I thought it was just me.

File "/usr/local/lib/python3.10/site-packages/transformers/pipelines/question_answering.py", line 270, in init
super().init(
File "/usr/local/lib/python3.10/site-packages/transformers/pipelines/base.py", line 939, in init
framework, model = infer_framework_load_model(model, config=model.config)

AttributeError: 'NoneType' object has no attribute 'config'

When reverting back to version 1.24.0, the problem goes away. I thought it was an opset issue and have been furiously testing with different fix/workarounds, but nothing. It occurs on multiple different onnx models.

I have tested backwards from 1.26.1, and this issue is only resolved at pip install --upgrade --upgrade-strategy eager "optimum[onnxruntime]==1.25.3"

something after 1.25.3 breaks the config / autoconfig..