BartForCausalLM analogs to ProphetNetForCausalLM (#9128)

* initiliaze bart4causalLM

* create BartDecoderWrapper, setters/getters

* delete spaces

* forward and additional methods

* update cache function, loss function, remove ngram* params in data class.

* add bartcausallm, bartdecoder testing

* correct bart for causal lm

* remove at

* add mbart as well

* up

* fix typo

* up

* correct

* add pegasusforcausallm

* add blenderbotforcausallm

* add blenderbotsmallforcausallm

* add marianforcausallm

* add test for MarianForCausalLM

* add Pegasus test

* add BlenderbotSmall test

* add blenderbot test

* fix a fail

* fix an import fail

* a fix

* fix

* Update modeling_pegasus.py

* fix models

* fix inputs_embeds setting getter

* adapt tests

* correct repo utils check

* finish test improvement

* fix tf models as well

* make style

* make fix-copies

* fix copies

* run all tests

* last changes

* fix all tests

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

tests/test_modeling_bart.py

@@ -27,7 +27,7 @@ from transformers.testing_utils import require_sentencepiece, require_tokenizers
 
 from .test_configuration_common import ConfigTester
 from .test_generation_utils import GenerationTesterMixin
-from .test_modeling_common import ModelTesterMixin, ids_tensor
+from .test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
 
 
 if is_torch_available():
@@ -36,6 +36,7 @@ if is_torch_available():
     from transformers import (
         AutoModelForSequenceClassification,
         BartConfig,
+        BartForCausalLM,
         BartForConditionalGeneration,
         BartForQuestionAnswering,
         BartForSequenceClassification,
@@ -178,7 +179,7 @@ class BartModelTester:
         self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
 
         # test that outputs are equal for slice
-        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
 
     def check_encoder_decoder_model_standalone(self, config, inputs_dict):
         model = BartModel(config=config).to(torch_device).eval()
@@ -719,3 +720,242 @@ class BartModelIntegrationTests(unittest.TestCase):
             hypotheses_batch.tolist(), clean_up_tokenization_spaces=True, skip_special_tokens=True
         )
         assert generated_summaries == EXPECTED
+
+
+class BartStandaloneDecoderModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        d_model=16,
+        decoder_seq_length=7,
+        is_training=True,
+        is_decoder=True,
+        use_attention_mask=True,
+        use_cache=False,
+        use_labels=True,
+        decoder_start_token_id=2,
+        decoder_ffn_dim=32,
+        decoder_layers=4,
+        encoder_attention_heads=4,
+        decoder_attention_heads=4,
+        max_position_embeddings=30,
+        is_encoder_decoder=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.hidden_size = d_model
+        self.num_hidden_layers = decoder_layers
+        self.decoder_layers = decoder_layers
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_attention_heads = decoder_attention_heads
+        self.num_attention_heads = decoder_attention_heads
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.is_encoder_decoder = is_encoder_decoder
+
+        self.scope = None
+        self.decoder_key_length = decoder_seq_length
+        self.base_model_out_len = 2
+        self.decoder_attention_idx = 1
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = BartConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.d_model,
+            encoder_layers=self.decoder_layers,
+            decoder_layers=self.decoder_layers,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            encoder_attention_heads=self.encoder_attention_heads,
+            decoder_attention_heads=self.decoder_attention_heads,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            use_cache=self.use_cache,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+            max_position_embeddings=self.max_position_embeddings,
+            is_encoder_decoder=self.is_encoder_decoder,
+        )
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        )
+
+    def prepare_config_and_inputs_for_decoder(self):
+        (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        ) = self.prepare_config_and_inputs()
+
+        encoder_hidden_states = floats_tensor([self.batch_size, self.decoder_seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            lm_labels,
+        )
+
+    def create_and_check_decoder_model_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        config.use_cache = True
+        model = BartDecoder(config=config).to(torch_device).eval()
+        # first forward pass
+        outputs = model(input_ids, use_cache=True)
+        outputs_use_cache_conf = model(input_ids)
+        outputs_no_past = model(input_ids, use_cache=False)
+
+        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
+        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
+
+        past_key_values = outputs["past_key_values"]
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+
+        output_from_no_past = model(next_input_ids)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def create_and_check_decoder_model_attention_mask_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        model = BartDecoder(config=config).to(torch_device).eval()
+
+        # create attention mask
+        attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+
+        half_seq_length = input_ids.shape[-1] // 2
+        attn_mask[:, half_seq_length:] = 0
+
+        # first forward pass
+        past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
+
+        # 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), dtype=torch.long, device=torch_device)],
+            dim=1,
+        )
+
+        # get two different outputs
+        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, attention_mask=attn_mask, past_key_values=past_key_values)[
+            "last_hidden_state"
+        ]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class BartStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (BartDecoder, BartForCausalLM) if is_torch_available() else ()
+    all_generative_model_classes = (BartForCausalLM,) if is_torch_available() else ()
+    test_pruning = False
+    is_encoder_decoder = False
+
+    def setUp(
+        self,
+    ):
+        self.model_tester = BartStandaloneDecoderModelTester(self, is_training=False)
+        self.config_tester = ConfigTester(self, config_class=BartConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_decoder_model_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
+
+    def test_decoder_model_attn_mask_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
+
+    def test_retain_grad_hidden_states_attentions(self):
+        # decoder cannot keep gradients
+        return

tests/test_modeling_blenderbot.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 """ Testing suite for the PyTorch Blenderbot model. """
 
-
 import tempfile
 import unittest
 
@@ -31,7 +30,11 @@ if is_torch_available():
     import torch
 
     from transformers import BlenderbotConfig, BlenderbotForConditionalGeneration, BlenderbotModel, BlenderbotTokenizer
-    from transformers.models.blenderbot.modeling_blenderbot import BlenderbotDecoder, BlenderbotEncoder
+    from transformers.models.blenderbot.modeling_blenderbot import (
+        BlenderbotDecoder,
+        BlenderbotEncoder,
+        BlenderbotForCausalLM,
+    )
 
 
 def prepare_blenderbot_inputs_dict(
@@ -165,7 +168,7 @@ class BlenderbotModelTester:
         self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
 
         # test that outputs are equal for slice
-        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
 
     def check_encoder_decoder_model_standalone(self, config, inputs_dict):
         model = BlenderbotModel(config=config).to(torch_device).eval()
@@ -300,3 +303,222 @@ class Blenderbot3BIntegrationTests(unittest.TestCase):
 
         assert "I think it's because we are so worried about what people think of us." == reply.strip()
         del model
+
+
+class BlenderbotStandaloneDecoderModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        d_model=16,
+        decoder_seq_length=7,
+        is_training=True,
+        is_decoder=True,
+        use_attention_mask=True,
+        use_cache=False,
+        use_labels=True,
+        decoder_start_token_id=2,
+        decoder_ffn_dim=32,
+        decoder_layers=4,
+        encoder_attention_heads=4,
+        decoder_attention_heads=4,
+        max_position_embeddings=30,
+        is_encoder_decoder=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.hidden_size = d_model
+        self.num_hidden_layers = decoder_layers
+        self.decoder_layers = decoder_layers
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_attention_heads = decoder_attention_heads
+        self.num_attention_heads = decoder_attention_heads
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.is_encoder_decoder = is_encoder_decoder
+
+        self.scope = None
+        self.decoder_key_length = decoder_seq_length
+        self.base_model_out_len = 2
+        self.decoder_attention_idx = 1
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = BlenderbotConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.d_model,
+            decoder_layers=self.decoder_layers,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            encoder_attention_heads=self.encoder_attention_heads,
+            decoder_attention_heads=self.decoder_attention_heads,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            use_cache=self.use_cache,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+            max_position_embeddings=self.max_position_embeddings,
+            is_encoder_decoder=self.is_encoder_decoder,
+        )
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        )
+
+    def create_and_check_decoder_model_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        config.use_cache = True
+        model = BlenderbotDecoder(config=config).to(torch_device).eval()
+        # first forward pass
+        outputs = model(input_ids, use_cache=True)
+        outputs_use_cache_conf = model(input_ids)
+        outputs_no_past = model(input_ids, use_cache=False)
+
+        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
+        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
+
+        past_key_values = outputs["past_key_values"]
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+
+        output_from_no_past = model(next_input_ids)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def create_and_check_decoder_model_attention_mask_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        model = BlenderbotDecoder(config=config).to(torch_device).eval()
+
+        # create attention mask
+        attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+
+        half_seq_length = input_ids.shape[-1] // 2
+        attn_mask[:, half_seq_length:] = 0
+
+        # first forward pass
+        past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
+        #        past_key_values = model(input_ids, use_cache=True)["past_key_values"]
+
+        # 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), dtype=torch.long, device=torch_device)],
+            dim=1,
+        )
+
+        # get two different outputs
+        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_values=past_key_values, attention_mask=attn_mask)[
+            "last_hidden_state"
+        ]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class BlenderbotStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (BlenderbotDecoder, BlenderbotForCausalLM) if is_torch_available() else ()
+    all_generative_model_classes = (BlenderbotForCausalLM,) if is_torch_available() else ()
+    test_pruning = False
+    is_encoder_decoder = False
+
+    def setUp(
+        self,
+    ):
+        self.model_tester = BlenderbotStandaloneDecoderModelTester(self, is_training=False)
+        self.config_tester = ConfigTester(self, config_class=BlenderbotConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_decoder_model_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
+
+    def test_decoder_model_attn_mask_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
+
+    def test_retain_grad_hidden_states_attentions(self):
+        # decoder cannot keep gradients
+        return

tests/test_modeling_blenderbot_small.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 """ Testing suite for the PyTorch BlenderbotSmall model. """
 
-
 import tempfile
 import unittest
 
@@ -39,6 +38,7 @@ if is_torch_available():
     from transformers.models.blenderbot_small.modeling_blenderbot_small import (
         BlenderbotSmallDecoder,
         BlenderbotSmallEncoder,
+        BlenderbotSmallForCausalLM,
     )
 
 
@@ -173,7 +173,7 @@ class BlenderbotSmallModelTester:
         self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
 
         # test that outputs are equal for slice
-        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
 
     def check_encoder_decoder_model_standalone(self, config, inputs_dict):
         model = BlenderbotSmallModel(config=config).to(torch_device).eval()
@@ -317,3 +317,221 @@ class Blenderbot90MIntegrationTests(unittest.TestCase):
             "have you ever been to a sam club? it's a great club in the south.",
             "have you ever heard of sam harris? he's an american singer, songwriter, and actor.",
         )
+
+
+class BlenderbotSmallStandaloneDecoderModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        d_model=16,
+        decoder_seq_length=7,
+        is_training=True,
+        is_decoder=True,
+        use_attention_mask=True,
+        use_cache=False,
+        use_labels=True,
+        decoder_start_token_id=2,
+        decoder_ffn_dim=32,
+        decoder_layers=4,
+        encoder_attention_heads=4,
+        decoder_attention_heads=4,
+        max_position_embeddings=30,
+        is_encoder_decoder=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.hidden_size = d_model
+        self.num_hidden_layers = decoder_layers
+        self.decoder_layers = decoder_layers
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_attention_heads = decoder_attention_heads
+        self.num_attention_heads = decoder_attention_heads
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.is_encoder_decoder = is_encoder_decoder
+
+        self.scope = None
+        self.decoder_key_length = decoder_seq_length
+        self.base_model_out_len = 2
+        self.decoder_attention_idx = 1
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = BlenderbotSmallConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.d_model,
+            decoder_layers=self.decoder_layers,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            encoder_attention_heads=self.encoder_attention_heads,
+            decoder_attention_heads=self.decoder_attention_heads,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            use_cache=self.use_cache,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+            max_position_embeddings=self.max_position_embeddings,
+            is_encoder_decoder=self.is_encoder_decoder,
+        )
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        )
+
+    def create_and_check_decoder_model_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        config.use_cache = True
+        model = BlenderbotSmallDecoder(config=config).to(torch_device).eval()
+        # first forward pass
+        outputs = model(input_ids, use_cache=True)
+        outputs_use_cache_conf = model(input_ids)
+        outputs_no_past = model(input_ids, use_cache=False)
+
+        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
+        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
+
+        past_key_values = outputs["past_key_values"]
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+
+        output_from_no_past = model(next_input_ids)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def create_and_check_decoder_model_attention_mask_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        model = BlenderbotSmallDecoder(config=config).to(torch_device).eval()
+
+        # create attention mask
+        attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+
+        half_seq_length = input_ids.shape[-1] // 2
+        attn_mask[:, half_seq_length:] = 0
+
+        # first forward pass
+        past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
+
+        # 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), dtype=torch.long, device=torch_device)],
+            dim=1,
+        )
+
+        # get two different outputs
+        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_values=past_key_values, attention_mask=attn_mask)[
+            "last_hidden_state"
+        ]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class BlenderbotSmallStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (BlenderbotSmallDecoder, BlenderbotSmallForCausalLM) if is_torch_available() else ()
+    all_generative_model_classes = (BlenderbotSmallForCausalLM,) if is_torch_available() else ()
+    test_pruning = False
+    is_encoder_decoder = False
+
+    def setUp(
+        self,
+    ):
+        self.model_tester = BlenderbotSmallStandaloneDecoderModelTester(self, is_training=False)
+        self.config_tester = ConfigTester(self, config_class=BlenderbotSmallConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_decoder_model_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
+
+    def test_decoder_model_attn_mask_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
+
+    def test_retain_grad_hidden_states_attentions(self):
+        # decoder cannot keep gradients
+        return

tests/test_modeling_encoder_decoder.py

@@ -20,6 +20,7 @@ import unittest
 from transformers import is_torch_available
 from transformers.testing_utils import require_torch, slow, torch_device
 
+from .test_modeling_bart import BartStandaloneDecoderModelTester
 from .test_modeling_bert import BertModelTester
 from .test_modeling_bert_generation import BertGenerationEncoderTester
 from .test_modeling_common import ids_tensor
@@ -34,6 +35,7 @@ if is_torch_available():
 
     from transformers import (
         AutoTokenizer,
+        BartForCausalLM,
         BertGenerationDecoder,
         BertGenerationEncoder,
         BertLMHeadModel,
@@ -828,3 +830,57 @@ class ProphetNetEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
 
     def test_encoder_decoder_model_shared_weights(self):
         pass
+
+
+@require_torch
+class BartEncoderDecoderModelTest(EncoderDecoderMixin, unittest.TestCase):
+    def get_encoder_decoder_model(self, config, decoder_config):
+        encoder_model = BertModel(config)
+        decoder_model = BartForCausalLM(decoder_config)
+        return encoder_model, decoder_model
+
+    def prepare_config_and_inputs(self):
+        model_tester_encoder = BertModelTester(self, batch_size=13)
+        model_tester_decoder = BartStandaloneDecoderModelTester(
+            self, batch_size=13, d_model=32, max_position_embeddings=512
+        )
+        encoder_config_and_inputs = model_tester_encoder.prepare_config_and_inputs()
+        decoder_config_and_inputs = model_tester_decoder.prepare_config_and_inputs_for_decoder()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = encoder_config_and_inputs
+        (
+            decoder_config,
+            decoder_input_ids,
+            decoder_attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            lm_labels,
+        ) = decoder_config_and_inputs
+
+        # make sure that cross attention layers are added
+        decoder_config.add_cross_attention = True
+        #  disable cache for now
+        decoder_config.use_cache = False
+        return {
+            "config": config,
+            "input_ids": input_ids,
+            "attention_mask": input_mask,
+            "decoder_config": decoder_config,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+            "encoder_hidden_states": encoder_hidden_states,
+            "labels": lm_labels,
+        }
+
+    def get_pretrained_model(self):
+        return EncoderDecoderModel.from_encoder_decoder_pretrained("bert-large-uncased", "facebook/bart-large")
+
+    def test_encoder_decoder_model_shared_weights(self):
+        pass

tests/test_modeling_marian.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 """ Testing suite for the PyTorch Marian model. """
 
-
 import tempfile
 import unittest
 
@@ -45,7 +44,12 @@ if is_torch_available():
         convert_hf_name_to_opus_name,
         convert_opus_name_to_hf_name,
     )
-    from transformers.models.marian.modeling_marian import MarianDecoder, MarianEncoder, shift_tokens_right
+    from transformers.models.marian.modeling_marian import (
+        MarianDecoder,
+        MarianEncoder,
+        MarianForCausalLM,
+        shift_tokens_right,
+    )
 
 
 def prepare_marian_inputs_dict(
@@ -182,7 +186,7 @@ class MarianModelTester:
         self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
 
         # test that outputs are equal for slice
-        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
 
     def check_encoder_decoder_model_standalone(self, config, inputs_dict):
         model = MarianModel(config=config).to(torch_device).eval()
@@ -546,3 +550,221 @@ class TestConversionUtils(unittest.TestCase):
             "en-de",
         ]
         self.assertListEqual(expected_opus_names, converted_opus_names)
+
+
+class MarianStandaloneDecoderModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        d_model=16,
+        decoder_seq_length=7,
+        is_training=True,
+        is_decoder=True,
+        use_attention_mask=True,
+        use_cache=False,
+        use_labels=True,
+        decoder_start_token_id=2,
+        decoder_ffn_dim=32,
+        decoder_layers=4,
+        encoder_attention_heads=4,
+        decoder_attention_heads=4,
+        max_position_embeddings=30,
+        is_encoder_decoder=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.hidden_size = d_model
+        self.num_hidden_layers = decoder_layers
+        self.decoder_layers = decoder_layers
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_attention_heads = decoder_attention_heads
+        self.num_attention_heads = decoder_attention_heads
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.is_encoder_decoder = is_encoder_decoder
+
+        self.scope = None
+        self.decoder_key_length = decoder_seq_length
+        self.base_model_out_len = 2
+        self.decoder_attention_idx = 1
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = MarianConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.d_model,
+            decoder_layers=self.decoder_layers,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            encoder_attention_heads=self.encoder_attention_heads,
+            decoder_attention_heads=self.decoder_attention_heads,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            use_cache=self.use_cache,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+            max_position_embeddings=self.max_position_embeddings,
+            is_encoder_decoder=self.is_encoder_decoder,
+        )
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        )
+
+    def create_and_check_decoder_model_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        config.use_cache = True
+        model = MarianDecoder(config=config).to(torch_device).eval()
+        # first forward pass
+        outputs = model(input_ids, use_cache=True)
+        outputs_use_cache_conf = model(input_ids)
+        outputs_no_past = model(input_ids, use_cache=False)
+
+        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
+        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
+
+        past_key_values = outputs["past_key_values"]
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+
+        output_from_no_past = model(next_input_ids)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def create_and_check_decoder_model_attention_mask_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        model = MarianDecoder(config=config).to(torch_device).eval()
+
+        # create attention mask
+        attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+
+        half_seq_length = input_ids.shape[-1] // 2
+        attn_mask[:, half_seq_length:] = 0
+
+        # first forward pass
+        past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
+
+        # 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), dtype=torch.long, device=torch_device)],
+            dim=1,
+        )
+
+        # get two different outputs
+        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, attention_mask=attn_mask, past_key_values=past_key_values)[
+            "last_hidden_state"
+        ]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class MarianStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (MarianDecoder, MarianForCausalLM) if is_torch_available() else ()
+    all_generative_model_classes = (MarianForCausalLM,) if is_torch_available() else ()
+    test_pruning = False
+    is_encoder_decoder = False
+
+    def setUp(
+        self,
+    ):
+        self.model_tester = MarianStandaloneDecoderModelTester(self, is_training=False)
+        self.config_tester = ConfigTester(self, config_class=MarianConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_decoder_model_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
+
+    def test_decoder_model_attn_mask_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
+
+    def test_retain_grad_hidden_states_attentions(self):
+        # decoder cannot keep gradients
+        return

tests/test_modeling_mbart.py

@@ -35,6 +35,7 @@ if is_torch_available():
         AutoTokenizer,
         BatchEncoding,
         MBartConfig,
+        MBartForCausalLM,
         MBartForConditionalGeneration,
         MBartForQuestionAnswering,
         MBartForSequenceClassification,
@@ -174,7 +175,7 @@ class MBartModelTester:
         self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
 
         # test that outputs are equal for slice
-        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
 
     def check_encoder_decoder_model_standalone(self, config, inputs_dict):
         model = MBartModel(config=config).to(torch_device).eval()
@@ -431,3 +432,221 @@ class MBartCC25IntegrationTest(AbstractSeq2SeqIntegrationTest):
             outputs, clean_up_tokenization_spaces=True, skip_special_tokens=True
         )[0]
         self.assertEqual(prediction, "of the best books I ever read!")
+
+
+class MBartStandaloneDecoderModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        d_model=16,
+        decoder_seq_length=7,
+        is_training=True,
+        is_decoder=True,
+        use_attention_mask=True,
+        use_cache=False,
+        use_labels=True,
+        decoder_start_token_id=2,
+        decoder_ffn_dim=32,
+        decoder_layers=4,
+        encoder_attention_heads=4,
+        decoder_attention_heads=4,
+        max_position_embeddings=30,
+        is_encoder_decoder=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.hidden_size = d_model
+        self.num_hidden_layers = decoder_layers
+        self.decoder_layers = decoder_layers
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_attention_heads = decoder_attention_heads
+        self.num_attention_heads = decoder_attention_heads
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.is_encoder_decoder = is_encoder_decoder
+
+        self.scope = None
+        self.decoder_key_length = decoder_seq_length
+        self.base_model_out_len = 2
+        self.decoder_attention_idx = 1
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = MBartConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.d_model,
+            decoder_layers=self.decoder_layers,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            encoder_attention_heads=self.encoder_attention_heads,
+            decoder_attention_heads=self.decoder_attention_heads,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            use_cache=self.use_cache,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+            max_position_embeddings=self.max_position_embeddings,
+            is_encoder_decoder=self.is_encoder_decoder,
+        )
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        )
+
+    def create_and_check_decoder_model_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        config.use_cache = True
+        model = MBartDecoder(config=config).to(torch_device).eval()
+        # first forward pass
+        outputs = model(input_ids, use_cache=True)
+        outputs_use_cache_conf = model(input_ids)
+        outputs_no_past = model(input_ids, use_cache=False)
+
+        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
+        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
+
+        past_key_values = outputs["past_key_values"]
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+
+        output_from_no_past = model(next_input_ids)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def create_and_check_decoder_model_attention_mask_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        model = MBartDecoder(config=config).to(torch_device).eval()
+
+        # create attention mask
+        attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+
+        half_seq_length = input_ids.shape[-1] // 2
+        attn_mask[:, half_seq_length:] = 0
+
+        # first forward pass
+        past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
+
+        # 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), dtype=torch.long, device=torch_device)],
+            dim=1,
+        )
+
+        # get two different outputs
+        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, attention_mask=attn_mask, past_key_values=past_key_values)[
+            "last_hidden_state"
+        ]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class MBartStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (MBartDecoder, MBartForCausalLM) if is_torch_available() else ()
+    all_generative_model_classes = (MBartForCausalLM,) if is_torch_available() else ()
+    test_pruning = False
+    is_encoder_decoder = False
+
+    def setUp(
+        self,
+    ):
+        self.model_tester = MBartStandaloneDecoderModelTester(self, is_training=False)
+        self.config_tester = ConfigTester(self, config_class=MBartConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_decoder_model_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
+
+    def test_decoder_model_attn_mask_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
+
+    def test_retain_grad_hidden_states_attentions(self):
+        # decoder cannot keep gradients
+        return

tests/test_modeling_pegasus.py

@@ -14,7 +14,6 @@
 # limitations under the License.
 """ Testing suite for the PyTorch PEGASUS model. """
 
-
 import tempfile
 import unittest
 
@@ -32,7 +31,7 @@ if is_torch_available():
     import torch
 
     from transformers import AutoModelForSeq2SeqLM, PegasusConfig, PegasusForConditionalGeneration, PegasusModel
-    from transformers.models.pegasus.modeling_pegasus import PegasusDecoder, PegasusEncoder
+    from transformers.models.pegasus.modeling_pegasus import PegasusDecoder, PegasusEncoder, PegasusForCausalLM
 
 
 def prepare_pegasus_inputs_dict(
@@ -166,7 +165,7 @@ class PegasusModelTester:
         self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
 
         # test that outputs are equal for slice
-        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
 
     def check_encoder_decoder_model_standalone(self, config, inputs_dict):
         model = PegasusModel(config=config).to(torch_device).eval()
@@ -308,3 +307,221 @@ class PegasusXSUMIntegrationTest(AbstractSeq2SeqIntegrationTest):
             "California's largest electricity provider has begun",
             "N-Dubz have revealed they were",
         ]
+
+
+class PegasusStandaloneDecoderModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        d_model=16,
+        decoder_seq_length=7,
+        is_training=True,
+        is_decoder=True,
+        use_attention_mask=True,
+        use_cache=False,
+        use_labels=True,
+        decoder_start_token_id=2,
+        decoder_ffn_dim=32,
+        decoder_layers=4,
+        encoder_attention_heads=4,
+        decoder_attention_heads=4,
+        max_position_embeddings=30,
+        is_encoder_decoder=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.hidden_size = d_model
+        self.num_hidden_layers = decoder_layers
+        self.decoder_layers = decoder_layers
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_attention_heads = decoder_attention_heads
+        self.num_attention_heads = decoder_attention_heads
+        self.eos_token_id = eos_token_id
+        self.bos_token_id = bos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.use_cache = use_cache
+        self.max_position_embeddings = max_position_embeddings
+        self.is_encoder_decoder = is_encoder_decoder
+
+        self.scope = None
+        self.decoder_key_length = decoder_seq_length
+        self.base_model_out_len = 2
+        self.decoder_attention_idx = 1
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = PegasusConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.d_model,
+            decoder_layers=self.decoder_layers,
+            decoder_ffn_dim=self.decoder_ffn_dim,
+            encoder_attention_heads=self.encoder_attention_heads,
+            decoder_attention_heads=self.decoder_attention_heads,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            use_cache=self.use_cache,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+            max_position_embeddings=self.max_position_embeddings,
+            is_encoder_decoder=self.is_encoder_decoder,
+        )
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        )
+
+    def create_and_check_decoder_model_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        config.use_cache = True
+        model = PegasusDecoder(config=config).to(torch_device).eval()
+        # first forward pass
+        outputs = model(input_ids, use_cache=True)
+        outputs_use_cache_conf = model(input_ids)
+        outputs_no_past = model(input_ids, use_cache=False)
+
+        self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
+        self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
+
+        past_key_values = outputs["past_key_values"]
+
+        # create hypothetical next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+
+        output_from_no_past = model(next_input_ids)["last_hidden_state"]
+        output_from_past = model(next_tokens, past_key_values=past_key_values)["last_hidden_state"]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def create_and_check_decoder_model_attention_mask_past(
+        self,
+        config,
+        input_ids,
+        attention_mask,
+        lm_labels,
+    ):
+        model = PegasusDecoder(config=config).to(torch_device).eval()
+
+        # create attention mask
+        attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
+
+        half_seq_length = input_ids.shape[-1] // 2
+        attn_mask[:, half_seq_length:] = 0
+
+        # first forward pass
+        past_key_values = model(input_ids, attention_mask=attn_mask, use_cache=True)["past_key_values"]
+
+        # 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), dtype=torch.long, device=torch_device)],
+            dim=1,
+        )
+
+        # get two different outputs
+        output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, attention_mask=attn_mask, past_key_values=past_key_values)[
+            "last_hidden_state"
+        ]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, next_input_ids.shape[-1] - 1, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
+
+        # test that outputs are equal for slice
+        assert torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3)
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            attention_mask,
+            lm_labels,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class PegasusStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (PegasusDecoder, PegasusForCausalLM) if is_torch_available() else ()
+    all_generative_model_classes = (PegasusForCausalLM,) if is_torch_available() else ()
+    test_pruning = False
+    is_encoder_decoder = False
+
+    def setUp(
+        self,
+    ):
+        self.model_tester = PegasusStandaloneDecoderModelTester(self, is_training=False)
+        self.config_tester = ConfigTester(self, config_class=PegasusConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_decoder_model_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_past(*config_and_inputs)
+
+    def test_decoder_model_attn_mask_past(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_attention_mask_past(*config_and_inputs)
+
+    def test_retain_grad_hidden_states_attentions(self):
+        # decoder cannot keep gradients
+        return