Adding gradient checkpointing to GPT2 (#7446)

* GPT2 gradient checkpointing

* find_unused_parameters removed if checkpointing

* find_unused_parameters removed if checkpointing

* Update src/transformers/configuration_gpt2.py

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Added a test for generation with checkpointing

* Update src/transformers/configuration_gpt2.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/configuration_gpt2.py

@@ -103,6 +103,8 @@ class GPT2Config(PretrainedConfig):
             :class:`~transformers.GPT2DoubleHeadsModel` and :class:`~transformers.TFGPT2DoubleHeadsModel`.
 
             The dropout ratio to be used after the projection and activation.
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.
 
     Example::
 
@@ -142,6 +144,7 @@ class GPT2Config(PretrainedConfig):
         summary_first_dropout=0.1,
         bos_token_id=50256,
         eos_token_id=50256,
+        gradient_checkpointing=False,
         **kwargs
     ):
         super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
@@ -164,6 +167,7 @@ class GPT2Config(PretrainedConfig):
         self.summary_activation = summary_activation
         self.summary_first_dropout = summary_first_dropout
         self.summary_proj_to_labels = summary_proj_to_labels
+        self.gradient_checkpointing = gradient_checkpointing
 
         self.bos_token_id = bos_token_id
         self.eos_token_id = eos_token_id

src/transformers/modeling_gpt2.py

@@ -15,7 +15,6 @@
 # limitations under the License.
 """PyTorch OpenAI GPT-2 model."""
 
-
 import os
 import warnings
 from dataclasses import dataclass
@@ -624,6 +623,25 @@ class GPT2Model(GPT2PreTrainedModel):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states.view(*output_shape),)
 
+            if getattr(self.config, "gradient_checkpointing", False):
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        # checkpointing only works with tuple returns, not with lists
+                        return tuple(output for output in module(*inputs, use_cache, output_attentions))
+
+                    return custom_forward
+
+                outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    layer_past,
+                    attention_mask,
+                    head_mask[i],
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
                 outputs = block(
                     hidden_states,
                     layer_past=layer_past,

src/transformers/trainer.py

@@ -679,8 +679,10 @@ class Trainer:
                 model,
                 device_ids=[self.args.local_rank],
                 output_device=self.args.local_rank,
-                find_unused_parameters=True,
+                find_unused_parameters=not getattr(model.config, "gradient_checkpointing", False),
             )
+        # find_unused_parameters breaks checkpointing as per
+        # https://github.com/huggingface/transformers/pull/4659#issuecomment-643356021
 
         if self.tb_writer is not None:
             self.tb_writer.add_text("args", self.args.to_json_string())

tests/test_modeling_gpt2.py

@@ -88,7 +88,7 @@ class GPT2ModelTester:
         self.bos_token_id = vocab_size - 1
         self.eos_token_id = vocab_size - 1
 
-    def prepare_config_and_inputs(self):
+    def prepare_config_and_inputs(self, gradient_checkpointing=False):
         input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
 
         input_mask = None
@@ -127,6 +127,7 @@ class GPT2ModelTester:
             bos_token_id=self.bos_token_id,
             eos_token_id=self.eos_token_id,
             return_dict=True,
+            gradient_checkpointing=gradient_checkpointing,
         )
 
         head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
@@ -269,6 +270,15 @@ class GPT2ModelTester:
         self.parent.assertEqual(result.loss.shape, ())
         self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
 
+    def create_and_check_forward_and_backwards(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+        model = GPT2LMHeadModel(config)
+        model.to(torch_device)
+
+        result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
+        self.parent.assertEqual(result.loss.shape, ())
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+        result.loss.backward()
+
     def create_and_check_double_lm_head_model(
         self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, *args
     ):
@@ -355,6 +365,10 @@ class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_double_lm_head_model(*config_and_inputs)
 
+    def test_gpt2_gradient_checkpointing(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs(gradient_checkpointing=True)
+        self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs)
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in GPT2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
@@ -366,7 +380,8 @@ class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
 class GPT2ModelLanguageGenerationTest(unittest.TestCase):
     @slow
     def test_lm_generate_gpt2(self):
-        model = GPT2LMHeadModel.from_pretrained("gpt2")
+        for checkpointing in [True, False]:
+            model = GPT2LMHeadModel.from_pretrained("gpt2", gradient_checkpointing=checkpointing)
             model.to(torch_device)
             input_ids = torch.tensor([[464, 3290]], dtype=torch.long, device=torch_device)  # The dog
             expected_output_ids = [