Fix gradient_checkpointing backward compatibility (#14408)

* Fix gradient_checkpointing backward compatibility

* Remove needless line

* make sure mask prob is big enough and length small enough

* Fix tests

Co-authored-by: patrickvonplaten <patrick.v.platen@gmail.com>

src/transformers/modeling_utils.py

@@ -412,6 +412,17 @@ class ModuleUtilsMixin:
         return 6 * self.estimate_tokens(input_dict) * self.num_parameters(exclude_embeddings=exclude_embeddings)
 
 
+def gradient_checkpointing_hook(module, _):
+    # Hook to enable backward compatibility for gradient checkpointing. Will be removed once all models have a
+    # proper post_init method.
+    if getattr(module.config, "gradient_checkpointing", False):
+        module.gradient_checkpointing_enable()
+        # Remove the attribute now that is has been consumed, so it's no saved in the config.
+        delattr(module.config, "gradient_checkpointing")
+    # The hook will remove itself after the first execution
+    module._gradient_checkpointing_hook.remove()
+
+
 class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin):
     r"""
     Base class for all models.
@@ -479,10 +490,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         # Save config and origin of the pretrained weights if given in model
         self.config = config
         self.name_or_path = config.name_or_path
-        if getattr(self.config, "gradient_checkpointing", False):
+        if self.supports_gradient_checkpointing:
-            self.gradient_checkpointing_enable()
+            self._gradient_checkpointing_hook = self.register_forward_pre_hook(gradient_checkpointing_hook)
-            # Remove the attribute now that is has been consumed, so it's no saved in the config.
-            delattr(self.config, "gradient_checkpointing")
 
     @classmethod
     def _from_config(cls, config, **kwargs):

src/transformers/models/detr/modeling_detr.py

@@ -783,7 +783,6 @@ class DetrClassificationHead(nn.Module):
 class DetrPreTrainedModel(PreTrainedModel):
     config_class = DetrConfig
     base_model_prefix = "model"
-    supports_gradient_checkpointing = True
 
     def _init_weights(self, module):
         std = self.config.init_std

src/transformers/models/layoutlmv2/modeling_layoutlmv2.py

@@ -504,7 +504,6 @@ class LayoutLMv2PreTrainedModel(PreTrainedModel):
     config_class = LayoutLMv2Config
     pretrained_model_archive_map = LAYOUTLMV2_PRETRAINED_MODEL_ARCHIVE_LIST
     base_model_prefix = "layoutlmv2"
-    supports_gradient_checkpointing = True
     _keys_to_ignore_on_load_missing = [r"position_ids"]
 
     def _init_weights(self, module):

tests/test_modeling_beit.py

@@ -239,6 +239,7 @@ class BeitModelTest(ModelTesterMixin, unittest.TestCase):
             if model_class.__name__ == "BeitForMaskedImageModeling":
                 continue
             model = model_class(config)
+            model.gradient_checkpointing_enable()
             model.to(torch_device)
             model.train()
             inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)

tests/test_modeling_common.py

@@ -209,6 +209,25 @@ class ModelTesterMixin:
                 )
                 self.assertTrue(len(load_result.unexpected_keys) == 0)
 
+    def test_gradient_checkpointing_backward_compatibility(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            if not model_class.supports_gradient_checkpointing:
+                continue
+
+            config.gradient_checkpointing = True
+            model = model_class(config)
+            # Model does not have gradient checkpointing activated yet, it will be done at the first forward.
+            self.assertFalse(model.is_gradient_checkpointing)
+
+            model.to(torch_device)
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+            _ = model(**inputs)
+
+            # Model has gradient checkpointing activated after the first forward.
+            self.assertTrue(model.is_gradient_checkpointing)
+
     def test_gradient_checkpointing_enable_disable(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
@@ -413,6 +432,7 @@ class ModelTesterMixin:
                 continue
             model = model_class(config)
             model.to(torch_device)
+            model.gradient_checkpointing_enable()
             model.train()
             inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
             loss = model(**inputs).loss

tests/test_modeling_deit.py

@@ -367,6 +367,7 @@ class DeiTModelTest(ModelTesterMixin, unittest.TestCase):
             if model_class.__name__ == "DeiTForImageClassificationWithTeacher":
                 continue
             model = model_class(config)
+            model.gradient_checkpointing_enable()
             model.to(torch_device)
             model.train()
             inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)

tests/test_modeling_unispeech_sat.py

@@ -65,6 +65,8 @@ class UniSpeechSatModelTester:
         layer_norm_eps=1e-5,
         hidden_act="gelu",
         initializer_range=0.02,
+        mask_time_prob=0.5,
+        mask_time_length=2,
         vocab_size=32,
         do_stable_layer_norm=False,
         scope=None,
@@ -92,6 +94,8 @@ class UniSpeechSatModelTester:
         self.initializer_range = initializer_range
         self.vocab_size = vocab_size
         self.do_stable_layer_norm = do_stable_layer_norm
+        self.mask_time_prob = mask_time_prob
+        self.mask_time_length = mask_time_length
         self.scope = scope
 
         output_seq_length = self.seq_length
@@ -120,6 +124,8 @@ class UniSpeechSatModelTester:
             conv_bias=self.conv_bias,
             num_conv_pos_embeddings=self.num_conv_pos_embeddings,
             num_conv_pos_embedding_groups=self.num_conv_pos_embedding_groups,
+            mask_time_prob=self.mask_time_prob,
+            mask_time_length=self.mask_time_length,
             num_hidden_layers=self.num_hidden_layers,
             num_attention_heads=self.num_attention_heads,
             hidden_dropout_prob=self.hidden_dropout_prob,

tests/test_modeling_wav2vec2.py

@@ -78,6 +78,8 @@ class Wav2Vec2ModelTester:
         layer_norm_eps=1e-5,
         hidden_act="gelu",
         initializer_range=0.02,
+        mask_time_prob=0.5,
+        mask_time_length=2,
         vocab_size=32,
         do_stable_layer_norm=False,
         scope=None,
@@ -105,6 +107,8 @@ class Wav2Vec2ModelTester:
         self.initializer_range = initializer_range
         self.vocab_size = vocab_size
         self.do_stable_layer_norm = do_stable_layer_norm
+        self.mask_time_prob = mask_time_prob
+        self.mask_time_length = mask_time_length
         self.scope = scope
 
         output_seq_length = self.seq_length
@@ -131,6 +135,8 @@ class Wav2Vec2ModelTester:
             conv_stride=self.conv_stride,
             conv_kernel=self.conv_kernel,
             conv_bias=self.conv_bias,
+            mask_time_prob=self.mask_time_prob,
+            mask_time_length=self.mask_time_length,
             num_conv_pos_embeddings=self.num_conv_pos_embeddings,
             num_conv_pos_embedding_groups=self.num_conv_pos_embedding_groups,
             num_hidden_layers=self.num_hidden_layers,