CI fails with dev dependencies: AttributeError: 'Qwen2_5_VLForConditionalGeneration' object has no attribute 'language_model'

[albertvillanova]

CI fails with dev dependencies: https://github.com/huggingface/trl/actions/runs/19767041653/job/56642242755

AttributeError: 'Qwen2_5_VLForConditionalGeneration' object has no attribute 'language_model'

FAILED tests/test_grpo_trainer.py::TestGRPOTrainer::test_training_vlm_and_liger[trl-internal-testing/tiny-Qwen2_5_VLForConditionalGeneration] - AttributeError: 'Qwen2_5_VLForConditionalGeneration' object has no attribute 'language_model'

Stacktrace:

   >       trainer = GRPOTrainer(
              model=model_id,
              reward_funcs=reward_func,
              args=training_args,
              train_dataset=dataset,
          )
  
  tests/test_grpo_trainer.py:1546: 
  _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
  trl/trainer/grpo_trainer.py:445: in __init__
      super().__init__(
  .venv/lib/python3.12/site-packages/transformers/trainer.py:486: in __init__
      _apply_liger_kernel_to_instance(model=model, **kernel_config)
  .venv/lib/python3.12/site-packages/liger_kernel/transformers/monkey_patch.py:2851: in _apply_liger_kernel_to_instance
      apply_fn(model=model, **applicable_kwargs)
  .venv/lib/python3.12/site-packages/liger_kernel/transformers/monkey_patch.py:1619: in apply_liger_kernel_to_qwen2_5_vl
      text_model: Qwen2_5_VLTextModel = model.language_model
                                        ^^^^^^^^^^^^^^^^^^^^
  _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
  
  self = Qwen2_5_VLForConditionalGeneration(
    (model): Qwen2_5_VLModel(
      (visual): Qwen2_5_VisionTransformerPretrainedModel...tary_emb): Qwen2_5_VLRotaryEmbedding()
      )
    )
    (lm_head): Linear(in_features=16, out_features=151936, bias=False)
  )
  name = 'language_model'
  
      def __getattr__(self, name: str) -> Union[Tensor, "Module"]:
          if "_parameters" in self.__dict__:
              _parameters = self.__dict__["_parameters"]
              if name in _parameters:
                  return _parameters[name]
          if "_buffers" in self.__dict__:
              _buffers = self.__dict__["_buffers"]
              if name in _buffers:
                  return _buffers[name]
          if "_modules" in self.__dict__:
              modules = self.__dict__["_modules"]
              if name in modules:
                  return modules[name]
  >       raise AttributeError(
              f"'{type(self).__name__}' object has no attribute '{name}'"
          )
  E       AttributeError: 'Qwen2_5_VLForConditionalGeneration' object has no attribute 'language_model'
  
  .venv/lib/python3.12/site-packages/torch/nn/modules/module.py:1964: AttributeError

[albertvillanova]

After investigation, it seems the issue appeared after the merge of this PR in transformers:

• 🚨 Generalize get_decoder() for multimodal and delete redundant code 🔪  transformers#42156

From the PR description, I see this breaking change is intended:

VLMs now will not have a property to get self.language_model directly from task-model and users will need to call self.get_decoder()

So I guess the liger-kernel library (one of our extra dependencies) should update their code and remove all the references to model.language_model, e.g.:

        if isinstance(model, (Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLModel)):
            # Note: language_model and visual properties can be accessed throught conditional class for BC.
            # Not sure if it is subject to changes in the future.
            # Reference: https://github.com/huggingface/transformers/blob/v4.52.4/src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py#L1823
            text_model: Qwen2_5_VLTextModel = model.language_model
            vision_model: Qwen2_5_VisionTransformerPretrainedModel = model.visual

I will track this upstream and look into a temporary workaround on our side until liger-kernel fixes it.

CC: @zucchini-nlp, who created the PR, for visibility.

[albertvillanova]

I opened an issue in liger-kernel repo:

• AttributeError for 'language_model' in transformers v5 linkedin/Liger-Kernel#960

[albertvillanova]

@zucchini-nlp, I have a question. If I understand correctly, now we should replace:

• model.language_model with model.get_decoder()
• model.visual with model.get_encoder(modality="image")

Is this right?

However, I think this is not backward compatible. To make the code backward compatible, we should replace:

• model.language_model with model.model.language_model
• model.visual with model.model.visual

as you did in tests/utils/test_modeling_utils.py: https://github.com/huggingface/transformers/pull/42156/files#diff-9828696cac704ee7096167deda48195c4d51892af221f7851f6404718796c5db

Is this right?

[zucchini-nlp]

However, I think this is not backward compatible

yep, the PR was breaking and I added the diff in migration guide. We do not want to have a property for multimodal models to skip the nested model.model

The second way in your comment will work only for qwen-vl, while the first way is supposed to be general for all models. Depending on use-case, feel free to use the one which suits better