[BUG] TransformersModel.from_model provides an incompletely initialized object compared to __init__

[huaanrui]

Describe the bug

When creating a TransformersModel instance using the from_model classmethod, the resulting object is missing several attributes (e.g.continuous_batching) that are normally initialized by the standard __init__ constructor. This leads to AttributeError exceptions during the evaluation pipeline, making the from_model method unreliable for integrating pre-loaded models.

To Reproduce

1, In an evaluation script, pre-load a model from transformers. 2. Instantiate the lighteval model wrapper using this pre-loaded model: model = TransformersModel.from_model(model=hf_model, ...) 3. Pass this model instance to a lighteval.pipeline.Pipeline. 4. Run pipeline.evaluate(). 5. The pipeline will crash with an AttributeError when an internal method attempts to access an attribute that was not initialized by from_model (for example, AttributeError: 'TransformersModel' object has no attribute 'continuous_batching').

Expected behavior

An object created via TransformersModel.from_model should be fully and correctly initialized, possessing the exact same set of attributes as an object created via the standard TransformersModel(config=...) constructor. It should be a drop-in replacement that works reliably within the pipeline.

Version info

Please provide your operating system, lighteval version or commit if you installed from main, and pip/conda environment if your problem concerns dependencies.

[NathanHB]

Hi ! thanks for raising this, I made a fix in #921. Can you test it on your side and make sure it fixes your issue ?