🚨Early-error🚨 config will error out if output_attentions=True and the attn implementation is wrong (#38288)

* Protect ParallelInterface

* early error out on output attention setting for no wraning in modeling

* modular update

* fixup

* update model tests

* update

* oups

* set model's config

* more cases

* ??

* properly fix

* fixup

* update

* last onces

* update

* fix?

* fix wrong merge commit

* fix hub test

* nits

* wow I am tired

* updates

* fix pipeline!

---------

Co-authored-by: Lysandre <[email protected]>

Author: ArthurZucker

docs/source/en/model_doc/jamba.md

@@ -99,7 +99,7 @@ quantization_config = BitsAndBytesConfig(load_in_8bit=True,
 device_map = {'model.embed_tokens': 0, 'model.layers.0': 0, 'model.layers.1': 0, 'model.layers.2': 0, 'model.layers.3': 0, 'model.layers.4': 0, 'model.layers.5': 0, 'model.layers.6': 0, 'model.layers.7': 0, 'model.layers.8': 0, 'model.layers.9': 1, 'model.layers.10': 1, 'model.layers.11': 1, 'model.layers.12': 1, 'model.layers.13': 1, 'model.layers.14': 1, 'model.layers.15': 1, 'model.layers.16': 1, 'model.layers.17': 1, 'model.layers.18': 2, 'model.layers.19': 2, 'model.layers.20': 2, 'model.layers.21': 2, 'model.layers.22': 2, 'model.layers.23': 2, 'model.layers.24': 2, 'model.layers.25': 2, 'model.layers.26': 2, 'model.layers.27': 3, 'model.layers.28': 3, 'model.layers.29': 3, 'model.layers.30': 3, 'model.layers.31': 3, 'model.layers.32': 3, 'model.layers.33': 3, 'model.layers.34': 3, 'model.layers.35': 3, 'model.layers.36': 4, 'model.layers.37': 4, 'model.layers.38': 4, 'model.layers.39': 4, 'model.layers.40': 4, 'model.layers.41': 4, 'model.layers.42': 4, 'model.layers.43': 4, 'model.layers.44': 4, 'model.layers.45': 5, 'model.layers.46': 5, 'model.layers.47': 5, 'model.layers.48': 5, 'model.layers.49': 5, 'model.layers.50': 5, 'model.layers.51': 5, 'model.layers.52': 5, 'model.layers.53': 5, 'model.layers.54': 6, 'model.layers.55': 6, 'model.layers.56': 6, 'model.layers.57': 6, 'model.layers.58': 6, 'model.layers.59': 6, 'model.layers.60': 6, 'model.layers.61': 6, 'model.layers.62': 6, 'model.layers.63': 7, 'model.layers.64': 7, 'model.layers.65': 7, 'model.layers.66': 7, 'model.layers.67': 7, 'model.layers.68': 7, 'model.layers.69': 7, 'model.layers.70': 7, 'model.layers.71': 7, 'model.final_layernorm': 7, 'lm_head': 7}
 model = AutoModelForCausalLM.from_pretrained("ai21labs/AI21-Jamba-Large-1.6",
                                              torch_dtype=torch.bfloat16,
-                                             attn_implementation="flash_attention_2",
+                    attn_implementation="flash_attention_2",
                                              quantization_config=quantization_config,
                                              device_map=device_map)
 

src/transformers/configuration_utils.py

@@ -214,7 +214,7 @@ def __init__(self, **kwargs):
         # Attributes with defaults
         self.return_dict = kwargs.pop("return_dict", True)
         self.output_hidden_states = kwargs.pop("output_hidden_states", False)
-        self.output_attentions = kwargs.pop("output_attentions", False)
+        self._output_attentions = kwargs.pop("output_attentions", False)
         self.torchscript = kwargs.pop("torchscript", False)  # Only used by PyTorch models
         self.torch_dtype = kwargs.pop("torch_dtype", None)  # Only used by PyTorch models
         self.use_bfloat16 = kwargs.pop("use_bfloat16", False)
@@ -331,6 +331,22 @@ def name_or_path(self) -> str:
     def name_or_path(self, value):
         self._name_or_path = str(value)  # Make sure that name_or_path is a string (for JSON encoding)
 
+    @property
+    def output_attentions(self):
+        """
+        `bool`: Whether or not the model should returns all attentions.
+        """
+        return self._output_attentions
+
+    @output_attentions.setter
+    def output_attentions(self, value):
+        if self._attn_implementation != "eager":
+            raise ValueError(
+                "The `output_attentions` attribute is not supported when using the `attn_implementation` set to "
+                f"{self._attn_implementation}. Please set it to 'eager' instead."
+            )
+        self._output_attentions = value
+
     @property
     def use_return_dict(self) -> bool:
         """
@@ -1004,6 +1020,8 @@ def _remove_keys_not_serialized(self, d: dict[str, Any]) -> None:
 
         if "_auto_class" in d:
             del d["_auto_class"]
+        if "_output_attentions" in d:
+            d["output_attentions"] = d.pop("_output_attentions")
         if "_commit_hash" in d:
             del d["_commit_hash"]
         if "_attn_implementation_internal" in d:

src/transformers/models/aria/modeling_aria.py

@@ -549,15 +549,8 @@ def forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         attention_interface: Callable = eager_attention_forward
-
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
-                logger.warning_once(
-                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
-                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-                )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
         attn_output, attn_weights = attention_interface(
             self,

src/transformers/models/bamba/modeling_bamba.py

@@ -313,15 +313,8 @@ def forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         attention_interface: Callable = eager_attention_forward
-
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
-                logger.warning_once(
-                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
-                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-                )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
         attn_output, attn_weights = attention_interface(
             self,

src/transformers/models/csm/modeling_csm.py

@@ -337,15 +337,8 @@ def forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         attention_interface: Callable = eager_attention_forward
-
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
-                logger.warning_once(
-                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
-                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-                )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
         attn_output, attn_weights = attention_interface(
             self,

src/transformers/models/emu3/modeling_emu3.py

@@ -206,15 +206,8 @@ def forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         attention_interface: Callable = eager_attention_forward
-
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
-                logger.warning_once(
-                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
-                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-                )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
         attn_output, attn_weights = attention_interface(
             self,

src/transformers/models/gemma/modeling_gemma.py

@@ -239,15 +239,8 @@ def forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         attention_interface: Callable = eager_attention_forward
-
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
-                logger.warning_once(
-                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
-                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-                )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
         attn_output, attn_weights = attention_interface(
             self,

src/transformers/models/glm/modeling_glm.py

@@ -201,15 +201,8 @@ def forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         attention_interface: Callable = eager_attention_forward
-
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
-                logger.warning_once(
-                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
-                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-                )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
         attn_output, attn_weights = attention_interface(
             self,

src/transformers/models/glm4/modeling_glm4.py

@@ -259,15 +259,8 @@ def forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         attention_interface: Callable = eager_attention_forward
-
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
-                logger.warning_once(
-                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
-                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-                )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
         attn_output, attn_weights = attention_interface(
             self,

src/transformers/models/granite/modeling_granite.py

@@ -165,15 +165,8 @@ def forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
         attention_interface: Callable = eager_attention_forward
-
         if self.config._attn_implementation != "eager":
-            if self.config._attn_implementation == "sdpa" and kwargs.get("output_attentions", False):
-                logger.warning_once(
-                    "`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to "
-                    'eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
-                )
-            else:
-                attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
 
         attn_output, attn_weights = attention_interface(
             self,