Fix attn mask gpt2 when using past (#3033)

* fix issue and add some tests

* fix issue and add some tests

* updated doc string gpt2

tests/test_modeling_gpt2.py

@@ -170,6 +170,72 @@ class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
             )
             self.parent.assertEqual(len(result["presents"]), config.n_layer)
 
+        def create_and_check_gpt2_model_past(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
+            model = GPT2Model(config=config)
+            model.to(torch_device)
+            model.eval()
+
+            # first forward pass
+            output, past = model(input_ids, token_type_ids=token_type_ids)
+
+            # create hypothetical next token and extent to next_input_ids
+            next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+            next_token_types = ids_tensor([self.batch_size, 1], self.type_vocab_size)
+
+            # append to next input_ids and token_type_ids
+            next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+            next_token_type_ids = torch.cat([token_type_ids, next_token_types], dim=-1)
+
+            output_from_no_past, _ = model(next_input_ids, token_type_ids=next_token_type_ids)
+            output_from_past, _ = model(next_tokens, token_type_ids=next_token_types, past=past)
+
+            # select random slice
+            random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+            output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
+            output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+            # test that outputs are equal for slice
+            self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
+        def create_and_check_gpt2_model_attention_mask_past(
+            self, config, input_ids, input_mask, head_mask, token_type_ids, *args
+        ):
+            model = GPT2Model(config=config)
+            model.to(torch_device)
+            model.eval()
+
+            # create attention mask
+            attn_mask = torch.ones(input_ids.shape).long()
+            half_seq_length = self.seq_length // 2
+            attn_mask[:, half_seq_length:] = 0
+
+            # first forward pass
+            output, past = model(input_ids, attention_mask=attn_mask)
+
+            # create hypothetical next token and extent to next_input_ids
+            next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+            # change a random masked slice from input_ids
+            random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
+            random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
+            input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
+
+            # append to next input_ids and attn_mask
+            next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+            attn_mask = torch.cat([attn_mask, torch.ones((attn_mask.shape[0], 1)).long()], dim=1)
+
+            # get two different outputs
+            output_from_no_past, _ = model(next_input_ids, attention_mask=attn_mask)
+            output_from_past, _ = model(next_tokens, past=past, attention_mask=attn_mask)
+
+            # select random slice
+            random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+            output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
+            output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+            # test that outputs are equal for slice
+            self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
         def create_and_check_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
             model = GPT2LMHeadModel(config)
             model.to(torch_device)
@@ -248,6 +314,14 @@ class GPT2ModelTest(ModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_gpt2_model(*config_and_inputs)
 
+    def test_gpt2_model_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_gpt2_model_past(*config_and_inputs)
+
+    def test_gpt2_model_att_mask_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_gpt2_model_attention_mask_past(*config_and_inputs)
+
     def test_gpt2_lm_head_model(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_lm_head_model(*config_and_inputs)