init

transformers/tests/models/mbart/test_modeling_mbart.py

@@ -0,0 +1,742 @@
+# Copyright 2021, The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Testing suite for the PyTorch MBART model."""
+
+import copy
+import tempfile
+import unittest
+from functools import cached_property
+
+from transformers import MBartConfig, is_torch_available
+from transformers.testing_utils import (
+    require_sentencepiece,
+    require_tokenizers,
+    require_torch,
+    require_torch_fp16,
+    slow,
+    torch_device,
+)
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        AutoTokenizer,
+        BatchEncoding,
+        MBartForCausalLM,
+        MBartForConditionalGeneration,
+        MBartForQuestionAnswering,
+        MBartForSequenceClassification,
+        MBartModel,
+    )
+    from transformers.models.mbart.modeling_mbart import MBartDecoder, MBartEncoder
+
+
+def prepare_mbart_inputs_dict(
+    config,
+    input_ids,
+    decoder_input_ids,
+    attention_mask=None,
+    decoder_attention_mask=None,
+):
+    if attention_mask is None:
+        attention_mask = input_ids.ne(config.pad_token_id)
+    if decoder_attention_mask is None:
+        decoder_attention_mask = decoder_input_ids.ne(config.pad_token_id)
+    return {
+        "input_ids": input_ids,
+        "decoder_input_ids": decoder_input_ids,
+        "attention_mask": attention_mask,
+        "decoder_attention_mask": attention_mask,
+    }
+
+
+class MBartModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_labels=False,
+        vocab_size=99,
+        hidden_size=16,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=4,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=100,
+        eos_token_id=2,
+        pad_token_id=1,
+        bos_token_id=0,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
+            3,
+        )
+        input_ids[:, -1] = self.eos_token_id  # Eos Token
+
+        decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        config = self.get_config()
+        inputs_dict = prepare_mbart_inputs_dict(config, input_ids, decoder_input_ids)
+        return config, inputs_dict
+
+    def get_config(self):
+        return MBartConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.hidden_size,
+            encoder_layers=self.num_hidden_layers,
+            decoder_layers=self.num_hidden_layers,
+            encoder_attention_heads=self.num_attention_heads,
+            decoder_attention_heads=self.num_attention_heads,
+            encoder_ffn_dim=self.intermediate_size,
+            decoder_ffn_dim=self.intermediate_size,
+            dropout=self.hidden_dropout_prob,
+            attention_dropout=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            pad_token_id=self.pad_token_id,
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config, inputs_dict = self.prepare_config_and_inputs()
+        return config, inputs_dict
+
+    def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
+        model = MBartModel(config=config).get_decoder().to(torch_device).eval()
+        input_ids = inputs_dict["input_ids"]
+        attention_mask = inputs_dict["attention_mask"]
+
+        # first forward pass
+        outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
+
+        output, past_key_values = outputs.to_tuple()
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_attn_mask = ids_tensor((self.batch_size, 3), 2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)
+
+        output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, attention_mask=next_attention_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[:, -3:, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
+
+        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-3))
+
+    def check_encoder_decoder_model_standalone(self, config, inputs_dict):
+        model = MBartModel(config=config).to(torch_device).eval()
+        outputs = model(**inputs_dict)
+
+        encoder_last_hidden_state = outputs.encoder_last_hidden_state
+        last_hidden_state = outputs.last_hidden_state
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            encoder = model.get_encoder()
+            encoder.save_pretrained(tmpdirname)
+            encoder = MBartEncoder.from_pretrained(tmpdirname).to(torch_device)
+
+        encoder_last_hidden_state_2 = encoder(inputs_dict["input_ids"], attention_mask=inputs_dict["attention_mask"])[
+            0
+        ]
+
+        self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max().item() < 1e-3)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            decoder = model.get_decoder()
+            decoder.save_pretrained(tmpdirname)
+            decoder = MBartDecoder.from_pretrained(tmpdirname).to(torch_device)
+
+        last_hidden_state_2 = decoder(
+            input_ids=inputs_dict["decoder_input_ids"],
+            attention_mask=inputs_dict["decoder_attention_mask"],
+            encoder_hidden_states=encoder_last_hidden_state,
+            encoder_attention_mask=inputs_dict["attention_mask"],
+        )[0]
+
+        self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max().item() < 1e-3)
+
+
+@require_torch
+class MBartModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (MBartModel, MBartForConditionalGeneration, MBartForSequenceClassification, MBartForQuestionAnswering)
+        if is_torch_available()
+        else ()
+    )
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": MBartModel,
+            "fill-mask": MBartForConditionalGeneration,
+            "question-answering": MBartForQuestionAnswering,
+            "summarization": MBartForConditionalGeneration,
+            "text-classification": MBartForSequenceClassification,
+            "text-generation": MBartForCausalLM,
+            "text2text-generation": MBartForConditionalGeneration,
+            "translation": MBartForConditionalGeneration,
+            "zero-shot": MBartForSequenceClassification,
+        }
+        if is_torch_available()
+        else {}
+    )
+    is_encoder_decoder = True
+    fx_compatible = False  # Fix me Michael
+    test_pruning = False
+    test_missing_keys = False
+
+    # TODO: Fix the failed tests
+    def is_pipeline_test_to_skip(
+        self,
+        pipeline_test_case_name,
+        config_class,
+        model_architecture,
+        tokenizer_name,
+        image_processor_name,
+        feature_extractor_name,
+        processor_name,
+    ):
+        if pipeline_test_case_name == "QAPipelineTests" and not tokenizer_name.endswith("Fast"):
+            return True
+
+        return False
+
+    def setUp(self):
+        self.model_tester = MBartModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=MBartConfig)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_save_load_strict(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs()
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+                model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
+            self.assertEqual(info["missing_keys"], [])
+
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def test_encoder_decoder_model_standalone(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_common()
+        self.model_tester.check_encoder_decoder_model_standalone(*config_and_inputs)
+
+    # MBartForSequenceClassification does not support inputs_embeds
+    def test_inputs_embeds(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in (MBartModel, MBartForConditionalGeneration, MBartForQuestionAnswering):
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
+
+            if not self.is_encoder_decoder:
+                input_ids = inputs["input_ids"]
+                del inputs["input_ids"]
+            else:
+                encoder_input_ids = inputs["input_ids"]
+                decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
+                del inputs["input_ids"]
+                inputs.pop("decoder_input_ids", None)
+
+            wte = model.get_input_embeddings()
+            if not self.is_encoder_decoder:
+                inputs["inputs_embeds"] = wte(input_ids)
+            else:
+                inputs["inputs_embeds"] = wte(encoder_input_ids)
+                inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
+
+            with torch.no_grad():
+                model(**inputs)[0]
+
+    @require_torch_fp16
+    def test_generate_fp16(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs()
+        input_ids = input_dict["input_ids"]
+        attention_mask = input_ids.ne(1).to(torch_device)
+        model = MBartForConditionalGeneration(config).eval().to(torch_device)
+        model.half()
+        model.generate(input_ids, attention_mask=attention_mask)
+        model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
+
+    def test_ensure_weights_are_shared(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs()
+
+        config.tie_word_embeddings = True
+        model = MBartForConditionalGeneration(config)
+
+        # MBart shares four weights.
+        # Not an issue to not have these correctly tied for torch.load, but it is an issue for safetensors.
+        self.assertEqual(
+            len(
+                {
+                    model.get_output_embeddings().weight.data_ptr(),
+                    model.get_input_embeddings().weight.data_ptr(),
+                    model.base_model.decoder.embed_tokens.weight.data_ptr(),
+                    model.base_model.encoder.embed_tokens.weight.data_ptr(),
+                }
+            ),
+            1,
+        )
+
+        config.tie_word_embeddings = False
+        model = MBartForConditionalGeneration(config)
+
+        # MBart shares four weights.
+        # Not an issue to not have these correctly tied for torch.load, but it is an issue for safetensors.
+        self.assertEqual(
+            len(
+                {
+                    model.get_output_embeddings().weight.data_ptr(),
+                    model.get_input_embeddings().weight.data_ptr(),
+                    model.base_model.decoder.embed_tokens.weight.data_ptr(),
+                    model.base_model.encoder.embed_tokens.weight.data_ptr(),
+                }
+            ),
+            2,
+        )
+
+    @unittest.skip(
+        reason="This architecture has tied weights by default and there is no way to remove it, check: https://github.com/huggingface/transformers/pull/31771#issuecomment-2210915245"
+    )
+    def test_load_save_without_tied_weights(self):
+        pass
+
+    def test_resize_embeddings_persists_embeddings_type(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs()
+
+        config.scale_embedding = True
+        model = MBartForConditionalGeneration(config)
+        old_type = type(model.model.decoder.embed_tokens)
+
+        model.resize_token_embeddings(new_num_tokens=config.vocab_size)
+
+        new_type = type(model.model.decoder.embed_tokens)
+        self.assertIs(old_type, new_type)
+
+
+def assert_tensors_close(a, b, atol=1e-12, prefix=""):
+    """If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error."""
+    if a is None and b is None:
+        return True
+    try:
+        if torch.allclose(a, b, atol=atol):
+            return True
+        raise
+    except Exception:
+        pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item()
+        if a.numel() > 100:
+            msg = f"tensor values are {pct_different:.1%} percent different."
+        else:
+            msg = f"{a} != {b}"
+        if prefix:
+            msg = prefix + ": " + msg
+        raise AssertionError(msg)
+
+
+def _long_tensor(tok_lst):
+    return torch.tensor(tok_lst, dtype=torch.long, device=torch_device)
+
+
+@require_torch
+@require_sentencepiece
+@require_tokenizers
+class AbstractSeq2SeqIntegrationTest(unittest.TestCase):
+    maxDiff = 1000  # longer string compare tracebacks
+    checkpoint_name = None
+
+    @classmethod
+    def setUpClass(cls):
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.checkpoint_name, use_fast=False)
+        return cls
+
+    @cached_property
+    def model(self):
+        """Only load the model if needed."""
+        model = MBartForConditionalGeneration.from_pretrained(self.checkpoint_name).to(torch_device)
+        if "cuda" in torch_device:
+            model = model.half()
+        return model
+
+
+@require_torch
+@require_sentencepiece
+@require_tokenizers
+class MBartEnroIntegrationTest(AbstractSeq2SeqIntegrationTest):
+    checkpoint_name = "facebook/mbart-large-en-ro"
+    src_text = [
+        " UN Chief Says There Is No Military Solution in Syria",
+        """ Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for Syria is that "there is no military solution" to the nearly five-year conflict and more weapons will only worsen the violence and misery for millions of people.""",
+    ]
+    tgt_text = [
+        "Şeful ONU declară că nu există o soluţie militară în Siria",
+        "Secretarul General Ban Ki-moon declară că răspunsul său la intensificarea sprijinului militar al Rusiei"
+        ' pentru Siria este că "nu există o soluţie militară" la conflictul de aproape cinci ani şi că noi arme nu vor'
+        " face decât să înrăutăţească violenţa şi mizeria pentru milioane de oameni.",
+    ]
+    expected_src_tokens = [8274, 127873, 25916, 7, 8622, 2071, 438, 67485, 53, 187895, 23, 51712, 2, 250004]
+
+    @slow
+    def test_enro_generate_one(self):
+        batch: BatchEncoding = self.tokenizer(
+            ["UN Chief Says There Is No Military Solution in Syria"], return_tensors="pt"
+        ).to(torch_device)
+        translated_tokens = self.model.generate(**batch)
+        decoded = self.tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
+        self.assertEqual(self.tgt_text[0], decoded[0])
+        # self.assertEqual(self.tgt_text[1], decoded[1])
+
+    @slow
+    def test_enro_generate_batch(self):
+        batch: BatchEncoding = self.tokenizer(self.src_text, return_tensors="pt", padding=True, truncation=True).to(
+            torch_device
+        )
+        translated_tokens = self.model.generate(**batch)
+        decoded = self.tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
+        assert self.tgt_text == decoded
+
+    def test_mbart_enro_config(self):
+        mbart_models = ["facebook/mbart-large-en-ro"]
+        expected = {"scale_embedding": True, "output_past": True}
+        for name in mbart_models:
+            config = MBartConfig.from_pretrained(name)
+            for k, v in expected.items():
+                try:
+                    self.assertEqual(v, getattr(config, k))
+                except AssertionError as e:
+                    e.args += (name, k)
+                    raise
+
+    def test_mbart_fast_forward(self):
+        config = MBartConfig(
+            vocab_size=99,
+            d_model=24,
+            encoder_layers=2,
+            decoder_layers=2,
+            encoder_attention_heads=2,
+            decoder_attention_heads=2,
+            encoder_ffn_dim=32,
+            decoder_ffn_dim=32,
+            max_position_embeddings=48,
+            add_final_layer_norm=True,
+        )
+        lm_model = MBartForConditionalGeneration(config).to(torch_device)
+        context = torch.tensor(
+            [[71, 82, 18, 33, 46, 91, 2], [68, 34, 26, 58, 30, 2, 1]], device=torch_device, dtype=torch.long
+        )
+        summary = torch.tensor([[82, 71, 82, 18, 2], [58, 68, 2, 1, 1]], device=torch_device, dtype=torch.long)
+        result = lm_model(input_ids=context, decoder_input_ids=summary, labels=summary)
+        expected_shape = (*summary.shape, config.vocab_size)
+        self.assertEqual(result.logits.shape, expected_shape)
+
+
+@require_torch
+@require_sentencepiece
+@require_tokenizers
+class MBartCC25IntegrationTest(AbstractSeq2SeqIntegrationTest):
+    checkpoint_name = "facebook/mbart-large-cc25"
+    src_text = [
+        " UN Chief Says There Is No Military Solution in Syria",
+        " I ate lunch twice yesterday",
+    ]
+    tgt_text = ["Şeful ONU declară că nu există o soluţie militară în Siria", "to be padded"]
+
+    @unittest.skip(reason="This test is broken, still generates english")
+    def test_cc25_generate(self):
+        inputs = self.tokenizer([self.src_text[0]], return_tensors="pt").to(torch_device)
+        translated_tokens = self.model.generate(
+            input_ids=inputs["input_ids"].to(torch_device),
+            decoder_start_token_id=self.tokenizer.lang_code_to_id["ro_RO"],
+        )
+        decoded = self.tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)
+        self.assertEqual(self.tgt_text[0], decoded[0])
+
+    @slow
+    def test_fill_mask(self):
+        inputs = self.tokenizer(["One of the best <mask> I ever read!"], return_tensors="pt").to(torch_device)
+        outputs = self.model.generate(
+            inputs["input_ids"], decoder_start_token_id=self.tokenizer.lang_code_to_id["en_XX"], num_beams=1
+        )
+        prediction: str = self.tokenizer.batch_decode(
+            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=2,
+        encoder_attention_heads=4,
+        decoder_attention_heads=4,
+        max_position_embeddings=50,
+        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,
+            num_hidden_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, use_cache=True
+        )["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 ()
+    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)
+
+    @unittest.skip(reason="Decoder cannot retain gradients")
+    def test_retain_grad_hidden_states_attentions(self):
+        return
+
+    @unittest.skip(reason="Decoder cannot retain gradients")
+    def test_flex_attention_with_grads(self):
+        return

transformers/tests/models/mbart/test_tokenization_mbart.py

@@ -0,0 +1,333 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import shutil
+import tempfile
+import unittest
+
+from transformers import SPIECE_UNDERLINE, BatchEncoding, MBartTokenizer, MBartTokenizerFast, is_torch_available
+from transformers.testing_utils import (
+    get_tests_dir,
+    nested_simplify,
+    require_sentencepiece,
+    require_tokenizers,
+    require_torch,
+)
+
+from ...test_tokenization_common import TokenizerTesterMixin
+
+
+SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
+
+
+if is_torch_available():
+    from transformers.models.mbart.modeling_mbart import shift_tokens_right
+
+EN_CODE = 250004
+RO_CODE = 250020
+
+
+@require_sentencepiece
+@require_tokenizers
+class MBartTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+    from_pretrained_id = "facebook/mbart-large-en-ro"
+    tokenizer_class = MBartTokenizer
+    rust_tokenizer_class = MBartTokenizerFast
+    test_rust_tokenizer = True
+    test_sentencepiece = True
+
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+
+        # We have a SentencePiece fixture for testing
+        tokenizer = MBartTokenizer(SAMPLE_VOCAB, keep_accents=True)
+        tokenizer.save_pretrained(cls.tmpdirname)
+
+    def test_full_tokenizer(self):
+        tokenizer = MBartTokenizer(SAMPLE_VOCAB, keep_accents=True)
+
+        tokens = tokenizer.tokenize("This is a test")
+        self.assertListEqual(tokens, ["▁This", "▁is", "▁a", "▁t", "est"])
+
+        self.assertListEqual(
+            tokenizer.convert_tokens_to_ids(tokens),
+            [value + tokenizer.fairseq_offset for value in [285, 46, 10, 170, 382]],
+        )
+
+        tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
+        self.assertListEqual(
+            tokens,
+            [
+                SPIECE_UNDERLINE + "I",
+                SPIECE_UNDERLINE + "was",
+                SPIECE_UNDERLINE + "b",
+                "or",
+                "n",
+                SPIECE_UNDERLINE + "in",
+                SPIECE_UNDERLINE + "",
+                "9",
+                "2",
+                "0",
+                "0",
+                "0",
+                ",",
+                SPIECE_UNDERLINE + "and",
+                SPIECE_UNDERLINE + "this",
+                SPIECE_UNDERLINE + "is",
+                SPIECE_UNDERLINE + "f",
+                "al",
+                "s",
+                "é",
+                ".",
+            ],
+        )
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertListEqual(
+            ids,
+            [
+                value + tokenizer.fairseq_offset
+                for value in [8, 21, 84, 55, 24, 19, 7, 2, 602, 347, 347, 347, 3, 12, 66, 46, 72, 80, 6, 2, 4]
+                #                                       ^ unk: 2 + 1 = 3                  unk: 2 + 1 = 3 ^
+            ],
+        )
+
+        back_tokens = tokenizer.convert_ids_to_tokens(ids)
+        self.assertListEqual(
+            back_tokens,
+            [
+                SPIECE_UNDERLINE + "I",
+                SPIECE_UNDERLINE + "was",
+                SPIECE_UNDERLINE + "b",
+                "or",
+                "n",
+                SPIECE_UNDERLINE + "in",
+                SPIECE_UNDERLINE + "",
+                "<unk>",
+                "2",
+                "0",
+                "0",
+                "0",
+                ",",
+                SPIECE_UNDERLINE + "and",
+                SPIECE_UNDERLINE + "this",
+                SPIECE_UNDERLINE + "is",
+                SPIECE_UNDERLINE + "f",
+                "al",
+                "s",
+                "<unk>",
+                ".",
+            ],
+        )
+
+    # overwrite from test_tokenization_common to speed up test
+    def test_save_pretrained(self):
+        if not self.test_slow_tokenizer:
+            # as we don't have a slow version, we can't compare the outputs between slow and fast versions
+            self.skipTest(reason="test_slow_tokenizer is set to False")
+
+        self.tokenizers_list[0] = (self.rust_tokenizer_class, "hf-internal-testing/tiny-random-mbart", {})
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                tokenizer_r = self.get_rust_tokenizer(pretrained_name, **kwargs)
+                tokenizer_p = self.get_tokenizer(pretrained_name, **kwargs)
+
+                tmpdirname2 = tempfile.mkdtemp()
+
+                tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2)
+                tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)
+
+                # Checks it save with the same files + the tokenizer.json file for the fast one
+                self.assertTrue(any("tokenizer.json" in f for f in tokenizer_r_files))
+                tokenizer_r_files = tuple(f for f in tokenizer_r_files if "tokenizer.json" not in f)
+                self.assertSequenceEqual(tokenizer_r_files, tokenizer_p_files)
+
+                # Checks everything loads correctly in the same way
+                tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
+                tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)
+
+                # Check special tokens are set accordingly on Rust and Python
+                for key in tokenizer_pp.special_tokens_map:
+                    self.assertTrue(hasattr(tokenizer_rp, key))
+                    # self.assertEqual(getattr(tokenizer_rp, key), getattr(tokenizer_pp, key))
+                    # self.assertEqual(getattr(tokenizer_rp, key + "_id"), getattr(tokenizer_pp, key + "_id"))
+
+                shutil.rmtree(tmpdirname2)
+
+                # Save tokenizer rust, legacy_format=True
+                tmpdirname2 = tempfile.mkdtemp()
+
+                tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2, legacy_format=True)
+                tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)
+
+                # Checks it save with the same files
+                self.assertSequenceEqual(tokenizer_r_files, tokenizer_p_files)
+
+                # Checks everything loads correctly in the same way
+                tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
+                tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)
+
+                # Check special tokens are set accordingly on Rust and Python
+                for key in tokenizer_pp.special_tokens_map:
+                    self.assertTrue(hasattr(tokenizer_rp, key))
+
+                shutil.rmtree(tmpdirname2)
+
+                # Save tokenizer rust, legacy_format=False
+                tmpdirname2 = tempfile.mkdtemp()
+
+                tokenizer_r_files = tokenizer_r.save_pretrained(tmpdirname2, legacy_format=False)
+                tokenizer_p_files = tokenizer_p.save_pretrained(tmpdirname2)
+
+                # Checks it saved the tokenizer.json file
+                self.assertTrue(any("tokenizer.json" in f for f in tokenizer_r_files))
+
+                # Checks everything loads correctly in the same way
+                tokenizer_rp = tokenizer_r.from_pretrained(tmpdirname2)
+                tokenizer_pp = tokenizer_p.from_pretrained(tmpdirname2)
+
+                # Check special tokens are set accordingly on Rust and Python
+                for key in tokenizer_pp.special_tokens_map:
+                    self.assertTrue(hasattr(tokenizer_rp, key))
+
+                shutil.rmtree(tmpdirname2)
+
+    @unittest.skip(reason="Need to fix this after #26538")
+    def test_training_new_tokenizer(self):
+        pass
+
+
+@require_torch
+@require_sentencepiece
+@require_tokenizers
+class MBartEnroIntegrationTest(unittest.TestCase):
+    checkpoint_name = "facebook/mbart-large-en-ro"
+    src_text = [
+        " UN Chief Says There Is No Military Solution in Syria",
+        """ Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for Syria is that "there is no military solution" to the nearly five-year conflict and more weapons will only worsen the violence and misery for millions of people.""",
+    ]
+    tgt_text = [
+        "Şeful ONU declară că nu există o soluţie militară în Siria",
+        "Secretarul General Ban Ki-moon declară că răspunsul său la intensificarea sprijinului militar al Rusiei"
+        ' pentru Siria este că "nu există o soluţie militară" la conflictul de aproape cinci ani şi că noi arme nu vor'
+        " face decât să înrăutăţească violenţele şi mizeria pentru milioane de oameni.",
+    ]
+    expected_src_tokens = [8274, 127873, 25916, 7, 8622, 2071, 438, 67485, 53, 187895, 23, 51712, 2, EN_CODE]
+
+    @classmethod
+    def setUpClass(cls):
+        cls.tokenizer: MBartTokenizer = MBartTokenizer.from_pretrained(
+            cls.checkpoint_name, src_lang="en_XX", tgt_lang="ro_RO"
+        )
+        cls.pad_token_id = 1
+        return cls
+
+    def check_language_codes(self):
+        self.assertEqual(self.tokenizer.fairseq_tokens_to_ids["ar_AR"], 250001)
+        self.assertEqual(self.tokenizer.fairseq_tokens_to_ids["en_EN"], 250004)
+        self.assertEqual(self.tokenizer.fairseq_tokens_to_ids["ro_RO"], 250020)
+
+    def test_enro_tokenizer_batch_encode_plus(self):
+        ids = self.tokenizer.batch_encode_plus(self.src_text).input_ids[0]
+        self.assertListEqual(self.expected_src_tokens, ids)
+
+    def test_enro_tokenizer_decode_ignores_language_codes(self):
+        self.assertIn(RO_CODE, self.tokenizer.all_special_ids)
+        generated_ids = [RO_CODE, 884, 9019, 96, 9, 916, 86792, 36, 18743, 15596, 5, 2]
+        result = self.tokenizer.decode(generated_ids, skip_special_tokens=True)
+        expected_romanian = self.tokenizer.decode(generated_ids[1:], skip_special_tokens=True)
+        self.assertEqual(result, expected_romanian)
+        self.assertNotIn(self.tokenizer.eos_token, result)
+
+    def test_enro_tokenizer_truncation(self):
+        src_text = ["this is gunna be a long sentence " * 20]
+        assert isinstance(src_text[0], str)
+        desired_max_length = 10
+        ids = self.tokenizer(src_text, max_length=desired_max_length, truncation=True).input_ids[0]
+        self.assertEqual(ids[-2], 2)
+        self.assertEqual(ids[-1], EN_CODE)
+        self.assertEqual(len(ids), desired_max_length)
+
+    def test_mask_token(self):
+        self.assertListEqual(self.tokenizer.convert_tokens_to_ids(["<mask>", "ar_AR"]), [250026, 250001])
+
+    def test_special_tokens_unaffacted_by_save_load(self):
+        tmpdirname = tempfile.mkdtemp()
+        original_special_tokens = self.tokenizer.fairseq_tokens_to_ids
+        self.tokenizer.save_pretrained(tmpdirname)
+        new_tok = MBartTokenizer.from_pretrained(tmpdirname)
+        self.assertDictEqual(new_tok.fairseq_tokens_to_ids, original_special_tokens)
+
+    @require_torch
+    def test_batch_fairseq_parity(self):
+        batch = self.tokenizer(self.src_text, text_target=self.tgt_text, padding=True, return_tensors="pt")
+        batch["decoder_input_ids"] = shift_tokens_right(batch["labels"], self.tokenizer.pad_token_id)
+
+        # fairseq batch: https://gist.github.com/sshleifer/cba08bc2109361a74ac3760a7e30e4f4
+        assert batch.input_ids[1][-2:].tolist() == [2, EN_CODE]
+        assert batch.decoder_input_ids[1][0].tolist() == RO_CODE
+        assert batch.decoder_input_ids[1][-1] == 2
+        assert batch.labels[1][-2:].tolist() == [2, RO_CODE]
+
+    @require_torch
+    def test_enro_tokenizer_prepare_batch(self):
+        batch = self.tokenizer(
+            self.src_text,
+            text_target=self.tgt_text,
+            padding=True,
+            truncation=True,
+            max_length=len(self.expected_src_tokens),
+            return_tensors="pt",
+        )
+
+        batch["decoder_input_ids"] = shift_tokens_right(batch["labels"], self.tokenizer.pad_token_id)
+
+        self.assertIsInstance(batch, BatchEncoding)
+
+        self.assertEqual((2, 14), batch.input_ids.shape)
+        self.assertEqual((2, 14), batch.attention_mask.shape)
+        result = batch.input_ids.tolist()[0]
+        self.assertListEqual(self.expected_src_tokens, result)
+        self.assertEqual(2, batch.decoder_input_ids[0, -1])  # EOS
+        # Test that special tokens are reset
+        self.assertEqual(self.tokenizer.prefix_tokens, [])
+        self.assertEqual(self.tokenizer.suffix_tokens, [self.tokenizer.eos_token_id, EN_CODE])
+
+    def test_seq2seq_max_length(self):
+        batch = self.tokenizer(self.src_text, padding=True, truncation=True, max_length=3, return_tensors="pt")
+        targets = self.tokenizer(
+            text_target=self.tgt_text, padding=True, truncation=True, max_length=10, return_tensors="pt"
+        )
+        labels = targets["input_ids"]
+        batch["decoder_input_ids"] = shift_tokens_right(labels, self.tokenizer.pad_token_id)
+
+        self.assertEqual(batch.input_ids.shape[1], 3)
+        self.assertEqual(batch.decoder_input_ids.shape[1], 10)
+
+    @require_torch
+    def test_tokenizer_translation(self):
+        inputs = self.tokenizer._build_translation_inputs(
+            "A test", return_tensors="pt", src_lang="en_XX", tgt_lang="ar_AR"
+        )
+
+        self.assertEqual(
+            nested_simplify(inputs),
+            {
+                # A, test, EOS, en_XX
+                "input_ids": [[62, 3034, 2, 250004]],
+                "attention_mask": [[1, 1, 1, 1]],
+                # ar_AR
+                "forced_bos_token_id": 250001,
+            },
+        )