CI fails with dev dependencies: AttributeError: 'Qwen3VLTextConfig' object has no attribute 'pad_token_id'

[albertvillanova]

https://github.com/huggingface/trl/actions/runs/21066816117/job/60586076404

AttributeError: 'Qwen3VLTextConfig' object has no attribute 'pad_token_id'. Did you mean: 'bos_token_id'?

  FAILED tests/test_sft_trainer.py::TestSFTTrainer::test_train_vlm[trl-internal-testing/tiny-Qwen3VLForConditionalGeneration] - AttributeError: 'Qwen3VLTextConfig' object has no attribute 'pad_token_id'. Did you mean: 'bos_token_id'?
  FAILED tests/test_utils.py::TestForwardMaskedLogits::test_vlm[trl-internal-testing/tiny-Qwen3VLForConditionalGeneration] - AttributeError: 'Qwen3VLTextConfig' object has no attribute 'pad_token_id'. Did you mean: 'bos_token_id'?

Stacktrace:

  >       model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="auto", device_map=device)
                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  
  tests/test_utils.py:927: 
  _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
  .venv/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py:372: in from_pretrained
      return model_class.from_pretrained(
  .venv/lib/python3.12/site-packages/transformers/modeling_utils.py:4020: in from_pretrained
      model = cls(config, *model_args, **model_kwargs)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  .venv/lib/python3.12/site-packages/transformers/models/qwen3_vl/modeling_qwen3_vl.py:1289: in __init__
      self.model = Qwen3VLModel(config)
                   ^^^^^^^^^^^^^^^^^^^^
  .venv/lib/python3.12/site-packages/transformers/models/qwen3_vl/modeling_qwen3_vl.py:938: in __init__
      self.language_model = Qwen3VLTextModel._from_config(config.text_config)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  .venv/lib/python3.12/site-packages/transformers/modeling_utils.py:1464: in _from_config
      model = cls(config, **kwargs)
              ^^^^^^^^^^^^^^^^^^^^^
  .venv/lib/python3.12/site-packages/transformers/models/qwen3_vl/modeling_qwen3_vl.py:807: in __init__
      self.padding_idx = config.pad_token_id
                         ^^^^^^^^^^^^^^^^^^^
  _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
  
  self = Qwen3VLTextConfig {
    "attention_bias": false,
    "attention_dropout": 0.0,
    "bos_token_id": 151643,
    "dtype": "bfloa...,
    "tie_word_embeddings": true,
    "transformers_version": "5.0.0.dev0",
    "use_cache": true,
    "vocab_size": 151936
  }
  
  key = 'pad_token_id'
  
      def __getattribute__(self, key):
          if key != "attribute_map" and key in super().__getattribute__("attribute_map"):
              key = super().__getattribute__("attribute_map")[key]
  >       return super().__getattribute__(key)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  E       AttributeError: 'Qwen3VLTextConfig' object has no attribute 'pad_token_id'. Did you mean: 'bos_token_id'?
  
  .venv/lib/python3.12/site-packages/transformers/configuration_utils.py:164: AttributeError

[albertvillanova]

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

• Untangle config inheritance transformers#41541

CC: @zucchini-nlp, who created the PR

[zucchini-nlp]

Thanks for reporting, indeed a text config needs to have a pad_token_id. Might have missed it in new models

Interesting that the slow CI didn't trigger any issues, let me also check if qwen3VL tests are being silently skipped on our side (update: there are no slow tests added at all 😢)

[qgallouedec]

@albertvillanova wanna xfail the tests in the meantime?

[albertvillanova]

Yes, I will open an issue upstream first, so I can reference it in the xfail message.