Add Aria

[aymeric-roucher]

What does this PR do?

Add rhymes-ai/Aria to transformers!

@ArthurZucker

[aymeric-roucher]

@ArthurZucker I have this error that we discussed in python utils/check_modular_conversion.py:

Traceback (most recent call last):
  File "/home/ubuntu/transformers/utils/check_modular_conversion.py", line 73, in <module>
    non_matching_files += compare_files(modular_file_path, args.fix_and_overwrite)
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/transformers/utils/check_modular_conversion.py", line 53, in compare_files
    generated_modeling_content = convert_modular_file(modular_file_path)
                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/transformers/utils/modular_model_converter.py", line 1138, in convert_modular_file
    wrapper.visit(cst_transformers)
  File "/home/ubuntu/venv/aria/lib/python3.12/site-packages/libcst/metadata/wrapper.py", line 204, in visit
    return self.module.visit(visitor)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/venv/aria/lib/python3.12/site-packages/libcst/_nodes/module.py", line 89, in visit
    result = super(Module, self).visit(visitor)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/venv/aria/lib/python3.12/site-packages/libcst/_nodes/base.py", line 236, in visit
    leave_result = visitor.on_leave(self, with_updated_children)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/venv/aria/lib/python3.12/site-packages/libcst/_visitors.py", line 71, in on_leave
    updated_node = leave_func(original_node, updated_node)
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ubuntu/transformers/utils/modular_model_converter.py", line 1111, in leave_Module
    self._recursively_add_all_new_needed_functions_in_files()
  File "/home/ubuntu/transformers/utils/modular_model_converter.py", line 1097, in _recursively_add_all_new_needed_functions_in_files
    dependency, body, self.all_definitions[dependency], parent=parent
                      ~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^
KeyError: 'select_best_resolution'

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

cc @Cyrilvallez as well!

[aymeric-roucher]

@ArthurZucker I added the second part of the check here to pass the test, it does not seem to create any missed imports so I let it but maybe it has unwanted consequences.

[ArthurZucker]

Nice then thanks!

[Cyrilvallez]

Ha nice @aymeric-roucher, indeed I forgot about built-ins when adding this! Thanks for correcting it! If you only check for if dependency in self.all_definitions, it should be enough though no? Built-ins will never be added to self.all_definitions during visit, so it would avoid having the large list

[aymeric-roucher]

It's done! ✅

[ArthurZucker]

LGTM, just

to make more readable / simplify

[DarkLight1337]

I'm unable to run the processor of this model in vLLM:

  File "/home/cyrus/miniconda3/envs/vllm/lib/python3.9/site-packages/transformers/models/aria/processing_aria.py", line 129, in __call__
    sample = sample.replace(self.tokenizer.image_token, self.tokenizer.image_token * num_crops)
  File "/home/cyrus/miniconda3/envs/vllm/lib/python3.9/site-packages/transformers/tokenization_utils_base.py", line 1108, in __getattr__
    raise AttributeError(f"{self.__class__.__name__} has no attribute {key}")
AttributeError: CachedLlamaTokenizerFast has no attribute image_token

Perhaps we need to add the image_token attribute to this processor, similar to the other VLMs?

[DarkLight1337]

This also happens with vanilla Transformers. Let me open a new issue on this.

Edit: Opened #35768