Improve @auto_docstring doc and rename args_doc.py to auto_docstring.py

docs/source/en/auto_docstring.md

@@ -64,9 +64,9 @@ Arguments can also be passed directly to `@auto_docstring` for more control. Use
     It builds upon the standard Transformer architecture with unique modifications.""",
     custom_args="""
     custom_parameter (`type`, *optional*, defaults to `default_value`):
-        A concise description for custom_parameter if not defined or overriding the description in `args_doc.py`.
+        A concise description for custom_parameter if not defined or overriding the description in `auto_docstring.py`.
     internal_helper_arg (`type`, *optional*, defaults to `default_value`):
-        A concise description for internal_helper_arg if not defined or overriding the description in `args_doc.py`.
+        A concise description for internal_helper_arg if not defined or overriding the description in `auto_docstring.py`.
     """
 )
 class MySpecialModel(PreTrainedModel):
@@ -85,13 +85,40 @@ class MySpecialModel(PreTrainedModel):
     def __init__(self, config: ConfigType, custom_parameter: "type" = "default_value", internal_helper_arg=None):
         r"""
         custom_parameter (`type`, *optional*, defaults to `default_value`):
-            A concise description for custom_parameter if not defined or overriding the description in `args_doc.py`.
+            A concise description for custom_parameter if not defined or overriding the description in `auto_docstring.py`.
         internal_helper_arg (`type`, *optional*, defaults to `default_value`):
-            A concise description for internal_helper_arg if not defined or overriding the description in `args_doc.py`.
+            A concise description for internal_helper_arg if not defined or overriding the description in `auto_docstring.py`.
         """
         # ...
 ```
 
+You should also use the `@auto_docstring` decorator for classes that inherit from [`~utils.ModelOutput`].
+
+```python
+@dataclass
+@auto_docstring(
+    custom_intro="""
+    Custom model outputs with additional fields.
+    """
+)
+class MyModelOutput(ImageClassifierOutput):
+    r"""
+    loss (`torch.FloatTensor`, *optional*):
+        The loss of the model.
+    custom_field (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*):
+        A custom output field specific to this model.
+    """
+
+    # Standard fields like hidden_states, logits, attentions etc. can be automatically documented if the description is the same as the standard arguments.
+    # However, given that the loss docstring is often different per model, you should document it in the docstring above.
+    loss: Optional[torch.FloatTensor] = None
+    logits: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
+    # Custom fields need to be documented in the docstring above
+    custom_field: Optional[torch.FloatTensor] = None
+```
+
 </hfoption>
 <hfoption id="functions">
 
@@ -171,7 +198,7 @@ class MyModel(PreTrainedModel):
 
 There are some rules for documenting different types of arguments and they're listed below.
 
-- Standard arguments (`input_ids`, `attention_mask`, `pixel_values`, etc.) are defined and retrieved from `args_doc.py`. It is the single source of truth for standard arguments and should not be redefined locally if an argument's description and shape is the same as an argument in `args_doc.py`.
+- Standard arguments (`input_ids`, `attention_mask`, `pixel_values`, etc.) are defined and retrieved from `auto_docstring.py`. It is the single source of truth for standard arguments and should not be redefined locally if an argument's description and shape is the same as an argument in `auto_docstring.py`.
 
     If a standard argument behaves differently in your model, then you can override it locally in a `r""" """` block. This local definition has a higher priority. For example, the `labels` argument is often customized per model and typically requires overriding.
 
@@ -245,15 +272,15 @@ When working with modular files (`modular_model.py`), follow the guidelines belo
 The `@auto_docstring` decorator automatically generates docstrings by:
 
 1. Inspecting the signature (arguments, types, defaults) of the decorated class' `__init__` method or the decorated function.
-2. Retrieving the predefined docstrings for common arguments (`input_ids`, `attention_mask`, etc.) from internal library sources like [`ModelArgs`], [`ImageProcessorArgs`], and the `args_doc.py` file.
+2. Retrieving the predefined docstrings for common arguments (`input_ids`, `attention_mask`, etc.) from internal library sources like [`ModelArgs`], [`ImageProcessorArgs`], and the `auto_docstring.py` file.
 3. Adding argument descriptions in one of two ways as shown below.
 
     | method | description | usage |
     |---|---|---|
     | `r""" """` | add custom docstring content directly to a method signature or within the `__init__` docstring | document new arguments or override standard descriptions |
     | `custom_args` | add custom docstrings for specific arguments directly in `@auto_docstring` | define docstring for new arguments once if they're repeated in multiple places in the modeling file |
 
-4. Adding class and function descriptions. For model classes with standard naming patterns, like `ModelForCausalLM`, or if it belongs to a pipeline, `@auto_docstring` automatically generates the appropriate descriptions with `ClassDocstring` from `args_doc.py`.
+4. Adding class and function descriptions. For model classes with standard naming patterns, like `ModelForCausalLM`, or if it belongs to a pipeline, `@auto_docstring` automatically generates the appropriate descriptions with `ClassDocstring` from `auto_docstring.py`.
 
     `@auto_docstring` also accepts the `custom_intro` argument to describe a class or function.
 

src/transformers/utils/__init__.py

@@ -21,7 +21,7 @@
 from packaging import version
 
 from .. import __version__
-from .args_doc import (
+from .auto_docstring import (
     ClassAttrs,
     ClassDocstring,
     ImageProcessorArgs,

src/transformers/utils/args_doc.py → src/transformers/utils/auto_docstring.py

@@ -1271,7 +1271,7 @@ def _get_model_info(func, parent_class):
             else:
                 config_class = "ModelConfig"
                 print(
-                    f"🚨 Config not found for {model_name_lowercase}. You can manually add it to HARDCODED_CONFIG_FOR_MODELS in utils/args_doc.py"
+                    f"🚨 Config not found for {model_name_lowercase}. You can manually add it to HARDCODED_CONFIG_FOR_MODELS in utils/auto_docstring.py"
                 )
 
     return model_name_lowercase, class_name, config_class
@@ -1893,27 +1893,150 @@ def auto_class_docstring(cls, custom_intro=None, custom_args=None, checkpoint=No
 
 
 def auto_docstring(obj=None, *, custom_intro=None, custom_args=None, checkpoint=None):
-    """
-    Automatically generates docstrings for classes and methods in the Transformers library.
-
-    This decorator can be used in the following forms:
-    @auto_docstring
-    def my_function(...):
-        ...
-    or
-    @auto_docstring()
-    def my_function(...):
-        ...
-    or
-    @auto_docstring(custom_intro="Custom intro", ...)
-    def my_function(...):
-        ...
+    r"""
+    Automatically generates comprehensive docstrings for model classes and methods in the Transformers library.
+
+    This decorator reduces boilerplate by automatically including standard argument descriptions while allowing
+    overrides to add new or custom arguments. It inspects function signatures, retrieves predefined docstrings
+    for common arguments (like `input_ids`, `attention_mask`, etc.), and generates complete documentation
+    including examples and return value descriptions.
+
+    For complete documentation and examples, read this [guide](https://huggingface.co/docs/transformers/auto_docstring).
+
+    Examples of usage:
+
+        Basic usage (no parameters):
+        ```python
+        @auto_docstring
+        class MyAwesomeModel(PreTrainedModel):
+            def __init__(self, config, custom_parameter: int = 10):
+                r'''
+                custom_parameter (`int`, *optional*, defaults to 10):
+                    Description of the custom parameter for MyAwesomeModel.
+                '''
+                super().__init__(config)
+                self.custom_parameter = custom_parameter
+        ```
+
+        Using `custom_intro` with a class:
+        ```python
+        @auto_docstring(
+            custom_intro="This model implements a novel attention mechanism for improved performance."
+        )
+        class MySpecialModel(PreTrainedModel):
+            def __init__(self, config, attention_type: str = "standard"):
+                r'''
+                attention_type (`str`, *optional*, defaults to "standard"):
+                    Type of attention mechanism to use.
+                '''
+                super().__init__(config)
+        ```
+
+        Using `custom_intro` with a method, and specify custom arguments and example directly in the docstring:
+        ```python
+        @auto_docstring(
+            custom_intro="Performs forward pass with enhanced attention computation."
+        )
+        def forward(
+            self,
+            input_ids: Optional[torch.Tensor] = None,
+            attention_mask: Optional[torch.Tensor] = None,
+        ):
+            r'''
+            custom_parameter (`int`, *optional*, defaults to 10):
+                Description of the custom parameter for MyAwesomeModel.
+
+            Example:
+
+            ```python
+            >>> model = MyAwesomeModel(config)
+            >>> model.forward(input_ids=torch.tensor([1, 2, 3]), attention_mask=torch.tensor([1, 1, 1]))
+            ```
+            '''
+        ```
+
+        Using `custom_args` to define reusable arguments:
+        ```python
+        VISION_ARGS = r'''
+        pixel_values (`torch.FloatTensor`, *optional*):
+            Pixel values of the input images.
+        image_features (`torch.FloatTensor`, *optional*):
+            Pre-computed image features for efficient processing.
+        '''
+
+        @auto_docstring(custom_args=VISION_ARGS)
+        def encode_images(self, pixel_values=None, image_features=None):
+            # ... method implementation
+        ```
+
+        Combining `custom_intro` and `custom_args`:
+        ```python
+        MULTIMODAL_ARGS = r'''
+        vision_features (`torch.FloatTensor`, *optional*):
+            Pre-extracted vision features from the vision encoder.
+        fusion_strategy (`str`, *optional*, defaults to "concat"):
+            Strategy for fusing text and vision modalities.
+        '''
+
+        @auto_docstring(
+            custom_intro="Processes multimodal inputs combining text and vision.",
+            custom_args=MULTIMODAL_ARGS
+        )
+        def forward(
+            self,
+            input_ids,
+            attention_mask=None,
+            vision_features=None,
+            fusion_strategy="concat"
+        ):
+            # ... multimodal processing
+        ```
+
+        Using with ModelOutput classes:
+        ```python
+        @dataclass
+        @auto_docstring(
+            custom_intro="Custom model outputs with additional fields."
+        )
+        class MyModelOutput(ImageClassifierOutput):
+            r'''
+            loss (`torch.FloatTensor`, *optional*):
+                The loss of the model.
+            custom_field (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*):
+                A custom output field specific to this model.
+            '''
+
+            # Standard fields like hidden_states, logits, attentions etc. can be automatically documented
+            # However, given that the loss docstring is often different per model, you should document it above
+            loss: Optional[torch.FloatTensor] = None
+            logits: Optional[torch.FloatTensor] = None
+            hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
+            attentions: Optional[tuple[torch.FloatTensor, ...]] = None
+            custom_field: Optional[torch.FloatTensor] = None
+        ```
 
     Args:
-        custom_intro (str, optional): Custom introduction text to add to the docstring. This will replace the default
-            introduction text generated by the decorator before the Args section.
-        checkpoint (str, optional): Checkpoint name to use in the docstring. This should be automatically inferred from the
-            model configuration class, but can be overridden if needed.
+        custom_intro (`str`, *optional*):
+            Custom introduction text to add to the docstring. This replaces the default
+            introduction text generated by the decorator before the Args section. Use this to describe what
+            makes your model or method special.
+        custom_args (`str`, *optional*):
+            Custom argument documentation in docstring format. This allows you to define
+            argument descriptions once and reuse them across multiple methods. The format should follow the
+            standard docstring convention: `arg_name (`type`, *optional*, defaults to `value`): Description.`
+        checkpoint (`str`, *optional*):
+            Checkpoint name to use in examples within the docstring. This is typically
+            automatically inferred from the model configuration class, but can be overridden if needed for
+            custom examples.
+
+    Note:
+        - Standard arguments (`input_ids`, `attention_mask`, `pixel_values`, etc.) are automatically documented
+          from predefined descriptions and should not be redefined unless their behavior differs in your model.
+        - New or custom arguments should be documented in the method's docstring using the `r''' '''` block
+          or passed via the `custom_args` parameter.
+        - For model classes, the decorator derives parameter descriptions from the `__init__` method's signature
+          and docstring.
+        - Return value documentation is automatically generated for methods that return ModelOutput subclasses.
     """
 
     def auto_docstring_decorator(obj):

utils/check_docstrings.py

@@ -49,7 +49,7 @@
 from git import Repo
 
 from transformers.utils import direct_transformers_import
-from transformers.utils.args_doc import (
+from transformers.utils.auto_docstring import (
     ImageProcessorArgs,
     ModelArgs,
     ModelOutputArgs,
@@ -1487,7 +1487,7 @@ def check_auto_docstrings(overwrite: bool = False, check_all: bool = False):
         if docstring_args_ro_remove_warnings:
             if not overwrite:
                 print(
-                    "Some docstrings are redundant with the ones in `args_doc.py` and will be removed. Run `make fix-copies` or `python utils/check_docstrings.py --fix_and_overwrite` to remove the redundant docstrings."
+                    "Some docstrings are redundant with the ones in `auto_docstring.py` and will be removed. Run `make fix-copies` or `python utils/check_docstrings.py --fix_and_overwrite` to remove the redundant docstrings."
                 )
             print(f"🚨 Redundant docstring for the following arguments in {candidate_file}:")
             for warning in docstring_args_ro_remove_warnings: