[core] Refactor of gradient_checkpointing

src/transformers/modeling_utils.py

@@ -14,6 +14,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import collections
+import functools
 import gc
 import importlib.metadata
 import inspect
@@ -1819,16 +1820,28 @@ def prune_heads(self, heads_to_prune: Dict[int, List[int]]):
 
         self.base_model._prune_heads(heads_to_prune)
 
-    def gradient_checkpointing_enable(self):
+    def gradient_checkpointing_enable(self, gradient_checkpointing_kwargs=None):
         """
         Activates gradient checkpointing for the current model.
 
         Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
         activations".
+
+        Args:
+            gradient_checkpointing_kwargs (dict, *optional*):
+                Additional keyword arguments passed along to the `torch.utils.checkpoint.checkpoint` function.
         """
         if not self.supports_gradient_checkpointing:
             raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
-        self.apply(partial(self._set_gradient_checkpointing, value=True))
+
+        if gradient_checkpointing_kwargs is None:
+            gradient_checkpointing_kwargs = {}
+
+        gradient_checkpointing_func = functools.partial(
+            torch.utils.checkpoint.checkpoint, **gradient_checkpointing_kwargs
+        )
+
+        self.apply(partial(self._set_gradient_checkpointing, gradient_checkpointing_func=gradient_checkpointing_func))
 
         if getattr(self, "_hf_peft_config_loaded", False):
             # When using PEFT + gradient checkpointing + Trainer we need to make sure the input has requires_grad=True
@@ -1845,7 +1858,7 @@ def gradient_checkpointing_disable(self):
         activations".
         """
         if self.supports_gradient_checkpointing:
-            self.apply(partial(self._set_gradient_checkpointing, value=False))
+            self.apply(partial(self._set_gradient_checkpointing, gradient_checkpointing_func=None))
 
         if getattr(self, "_hf_peft_config_loaded", False):
             self.disable_input_require_grads()

src/transformers/models/align/modeling_align.py

@@ -1095,20 +1095,15 @@ def forward(
             past_key_value = past_key_values[i] if past_key_values is not None else None
 
             if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, past_key_value, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
+                layer_outputs = self.gradient_checkpointing_func(
+                    layer_module.forward,
                     hidden_states,
                     attention_mask,
                     layer_head_mask,
                     encoder_hidden_states,
                     encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
                 )
             else:
                 layer_outputs = layer_module(
@@ -1197,9 +1192,10 @@ def _init_weights(self, module):
             module.bias.data.zero_()
             module.weight.data.fill_(1.0)
 
-    def _set_gradient_checkpointing(self, module, value=False):
+    def _set_gradient_checkpointing(self, module, gradient_checkpointing_func=None):
         if isinstance(module, (AlignTextModel, AlignVisionModel)):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
 
 
 @add_start_docstrings(

src/transformers/models/altclip/modeling_altclip.py

@@ -646,20 +646,15 @@ def forward(
             past_key_value = past_key_values[i] if past_key_values is not None else None
 
             if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, past_key_value, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
+                layer_outputs = self.gradient_checkpointing_func(
+                    layer_module.forward,
                     hidden_states,
                     attention_mask,
                     layer_head_mask,
                     encoder_hidden_states,
                     encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
                 )
             else:
                 layer_outputs = layer_module(
@@ -960,18 +955,12 @@ def forward(
             if output_hidden_states:
                 encoder_states = encoder_states + (hidden_states,)
             if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(encoder_layer),
+                layer_outputs = self.gradient_checkpointing_func(
+                    encoder_layer.forward,
                     hidden_states,
                     attention_mask,
                     causal_attention_mask,
+                    output_attentions,
                 )
             else:
                 layer_outputs = encoder_layer(
@@ -1089,11 +1078,13 @@ def _init_weights(self, module):
             if module.padding_idx is not None:
                 module.weight.data[module.padding_idx].zero_()
 
-    def _set_gradient_checkpointing(self, module, value=False):
+    def _set_gradient_checkpointing(self, module, gradient_checkpointing_func=None):
         if isinstance(module, AltCLIPEncoder):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
         if isinstance(module, AltRobertaEncoder):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
 
 
 # Copied from transformers.models.clip.modeling_clip.CLIPVisionTransformer with CLIPVisionTransformer->AltCLIPVisionTransformer,CLIPVisionConfig->AltCLIPVisionConfig,CLIPVisionEmbeddings->AltCLIPVisionEmbeddings,CLIPEncoder->AltCLIPEncoder,CLIP_VISION_INPUTS_DOCSTRING->ALTCLIP_VISION_INPUTS_DOCSTRING

src/transformers/models/audio_spectrogram_transformer/modeling_audio_spectrogram_transformer.py

@@ -336,17 +336,11 @@ def forward(
             layer_head_mask = head_mask[i] if head_mask is not None else None
 
             if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
+                layer_outputs = self.gradient_checkpointing_func(
+                    layer_module.forward,
                     hidden_states,
                     layer_head_mask,
+                    output_attentions,
                 )
             else:
                 layer_outputs = layer_module(hidden_states, layer_head_mask, output_attentions)
@@ -395,9 +389,10 @@ def _init_weights(self, module: Union[nn.Linear, nn.Conv2d, nn.LayerNorm]) -> No
             module.weight.data.fill_(1.0)
 
     # Copied from transformers.models.vit.modeling_vit.ViTPreTrainedModel._set_gradient_checkpointing with ViT->AST
-    def _set_gradient_checkpointing(self, module: ASTEncoder, value: bool = False) -> None:
+    def _set_gradient_checkpointing(self, module: ASTEncoder, gradient_checkpointing_func=None) -> None:
         if isinstance(module, ASTEncoder):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
 
 
 AUDIO_SPECTROGRAM_TRANSFORMER_START_DOCSTRING = r"""

src/transformers/models/autoformer/modeling_autoformer.py

@@ -946,9 +946,10 @@ def _init_weights(self, module):
             if module.padding_idx is not None:
                 module.weight.data[module.padding_idx].zero_()
 
-    def _set_gradient_checkpointing(self, module, value=False):
+    def _set_gradient_checkpointing(self, module, gradient_checkpointing_func=None):
         if isinstance(module, (AutoformerDecoder, AutoformerEncoder)):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
 
 
 AUTOFORMER_START_DOCSTRING = r"""
@@ -1207,18 +1208,12 @@ def forward(
                 layer_outputs = (None, None)
             else:
                 if self.gradient_checkpointing and self.training:
-
-                    def create_custom_forward(module):
-                        def custom_forward(*inputs):
-                            return module(*inputs, output_attentions)
-
-                        return custom_forward
-
-                    layer_outputs = torch.utils.checkpoint.checkpoint(
-                        create_custom_forward(encoder_layer),
+                    layer_outputs = self.gradient_checkpointing_func(
+                        encoder_layer.forward,
                         hidden_states,
                         attention_mask,
                         (head_mask[idx] if head_mask is not None else None),
+                        output_attentions,
                     )
                 else:
                     layer_outputs = encoder_layer(
@@ -1433,8 +1428,8 @@ def custom_forward(*inputs):
 
                     return custom_forward
 
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(decoder_layer),
+                layer_outputs = self.gradient_checkpointing_func(
+                    decoder_layer.forward,
                     hidden_states,
                     attention_mask,
                     encoder_hidden_states,

src/transformers/models/bark/modeling_bark.py

@@ -313,9 +313,10 @@ def device(self) -> torch.device:
 
         return get_parameter_device(self)
 
-    def _set_gradient_checkpointing(self, module, value=False):
+    def _set_gradient_checkpointing(self, module, gradient_checkpointing_func=None):
         if isinstance(module, BarkCausalModel) or isinstance(module, BarkFineModel) or isinstance(module, BarkModel):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
 
 
 BARK_MODEL_START_DOCSTRING = """
@@ -637,20 +638,14 @@ def forward(
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        # None for past_key_value
-                        return module(*inputs, use_cache, output_attentions)
-
-                    return custom_forward
-
-                outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(block),
+                outputs = self.gradient_checkpointing_func(
+                    block.forward,
                     hidden_states,
                     None,
                     attention_mask,
                     head_mask[i],
+                    use_cache,
+                    output_attentions,
                 )
             else:
                 outputs = block(

src/transformers/models/bart/modeling_bart.py

@@ -521,9 +521,10 @@ def _init_weights(self, module):
             if module.padding_idx is not None:
                 module.weight.data[module.padding_idx].zero_()
 
-    def _set_gradient_checkpointing(self, module, value=False):
+    def _set_gradient_checkpointing(self, module, gradient_checkpointing_func=None):
         if isinstance(module, (BartDecoder, BartEncoder)):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
 
     @property
     def dummy_inputs(self):
@@ -854,18 +855,12 @@ def forward(
                 layer_outputs = (None, None)
             else:
                 if self.gradient_checkpointing and self.training:
-
-                    def create_custom_forward(module):
-                        def custom_forward(*inputs):
-                            return module(*inputs, output_attentions)
-
-                        return custom_forward
-
-                    layer_outputs = torch.utils.checkpoint.checkpoint(
-                        create_custom_forward(encoder_layer),
+                    layer_outputs = self.gradient_checkpointing_func(
+                        encoder_layer.forward,
                         hidden_states,
                         attention_mask,
                         (head_mask[idx] if head_mask is not None else None),
+                        output_attentions,
                     )
                 else:
                     layer_outputs = encoder_layer(
@@ -1118,8 +1113,8 @@ def custom_forward(*inputs):
 
                     return custom_forward
 
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(decoder_layer),
+                layer_outputs = self.gradient_checkpointing_func(
+                    decoder_layer.forward,
                     hidden_states,
                     attention_mask,
                     encoder_hidden_states,

src/transformers/models/beit/modeling_beit.py

@@ -510,17 +510,11 @@ def forward(
             layer_head_mask = head_mask[i] if head_mask is not None else None
 
             if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
+                layer_outputs = self.gradient_checkpointing_func(
+                    layer_module.forward,
                     hidden_states,
                     layer_head_mask,
+                    output_attentions,
                 )
             else:
                 relative_position_bias = (
@@ -572,9 +566,10 @@ def _init_weights(self, module):
             module.bias.data.zero_()
             module.weight.data.fill_(1.0)
 
-    def _set_gradient_checkpointing(self, module, value=False):
+    def _set_gradient_checkpointing(self, module, gradient_checkpointing_func=None):
         if isinstance(module, BeitEncoder):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
 
 
 BEIT_START_DOCSTRING = r"""

src/transformers/models/bert/modeling_bert.py

@@ -593,20 +593,15 @@ def forward(
             past_key_value = past_key_values[i] if past_key_values is not None else None
 
             if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        return module(*inputs, past_key_value, output_attentions)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(layer_module),
+                layer_outputs = self.gradient_checkpointing_func(
+                    layer_module.forward,
                     hidden_states,
                     attention_mask,
                     layer_head_mask,
                     encoder_hidden_states,
                     encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
                 )
             else:
                 layer_outputs = layer_module(
@@ -762,9 +757,10 @@ def _init_weights(self, module):
             module.bias.data.zero_()
             module.weight.data.fill_(1.0)
 
-    def _set_gradient_checkpointing(self, module, value=False):
+    def _set_gradient_checkpointing(self, module, gradient_checkpointing_func=None):
         if isinstance(module, BertEncoder):
-            module.gradient_checkpointing = value
+            module.gradient_checkpointing_func = gradient_checkpointing_func
+            module.gradient_checkpointing = gradient_checkpointing_func is not None
 
 
 @dataclass