init

transformers/tests/models/roc_bert/test_modeling_roc_bert.py

@@ -0,0 +1,859 @@
+# Copyright 2022 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 RoCBert model."""
+
+import inspect
+import tempfile
+import unittest
+
+from transformers import RoCBertConfig, is_torch_available
+from transformers.models.auto import get_values
+from transformers.testing_utils import require_torch, slow, torch_device
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+from ...test_pipeline_mixin import PipelineTesterMixin
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        MODEL_FOR_PRETRAINING_MAPPING,
+        DataCollatorWithFlattening,
+        RoCBertForCausalLM,
+        RoCBertForMaskedLM,
+        RoCBertForMultipleChoice,
+        RoCBertForPreTraining,
+        RoCBertForQuestionAnswering,
+        RoCBertForSequenceClassification,
+        RoCBertForTokenClassification,
+        RoCBertModel,
+    )
+
+
+class RoCBertModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        pronunciation_vocab_size=99,
+        shape_vocab_size=99,
+        pronunciation_embed_dim=32,
+        shape_embed_dim=32,
+        hidden_size=32,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.pronunciation_vocab_size = pronunciation_vocab_size
+        self.shape_vocab_size = shape_vocab_size
+        self.pronunciation_embed_dim = pronunciation_embed_dim
+        self.shape_embed_dim = shape_embed_dim
+        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.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+        input_shape_ids = ids_tensor([self.batch_size, self.seq_length], self.shape_vocab_size)
+        input_pronunciation_ids = ids_tensor([self.batch_size, self.seq_length], self.pronunciation_vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config()
+
+        return (
+            config,
+            input_ids,
+            input_shape_ids,
+            input_pronunciation_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        )
+
+    def get_config(self):
+        return RoCBertConfig(
+            vocab_size=self.vocab_size,
+            shape_vocab_size=self.shape_vocab_size,
+            pronunciation_vocab_size=self.pronunciation_vocab_size,
+            shape_embed_dim=self.shape_embed_dim,
+            pronunciation_embed_dim=self.pronunciation_embed_dim,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+    def prepare_config_and_inputs_for_decoder(self):
+        (
+            config,
+            input_ids,
+            input_shape_ids,
+            input_pronunciation_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = self.prepare_config_and_inputs()
+
+        config.is_decoder = True
+        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            input_shape_ids,
+            input_pronunciation_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+    def create_and_check_model(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        model = RoCBertModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+        )
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            token_type_ids=token_type_ids,
+        )
+        result = model(input_ids, input_shape_ids=input_shape_ids, input_pronunciation_ids=input_pronunciation_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_model_as_decoder(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.add_cross_attention = True
+        model = RoCBertModel(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+        )
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+        )
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+
+    def create_and_check_for_masked_lm(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        model = RoCBertForMaskedLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            labels=token_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.is_decoder = True
+        config.add_cross_attention = True
+        model = RoCBertForCausalLM(config=config)
+        model.to(torch_device)
+        model.eval()
+
+        # first forward pass
+        outputs = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_shape_tokens = ids_tensor((self.batch_size, 3), config.shape_vocab_size)
+        next_pronunciation_tokens = ids_tensor((self.batch_size, 3), config.pronunciation_vocab_size)
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_input_shape_ids = torch.cat([input_shape_ids, next_shape_tokens], dim=-1)
+        next_input_pronunciation_ids = torch.cat([input_pronunciation_ids, next_pronunciation_tokens], dim=-1)
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            input_shape_ids=next_input_shape_ids,
+            input_pronunciation_ids=next_input_pronunciation_ids,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+        output_from_past = model(
+            next_tokens,
+            input_shape_ids=next_shape_tokens,
+            input_pronunciation_ids=next_pronunciation_tokens,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+
+        # 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 create_and_check_for_question_answering(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        model = RoCBertForQuestionAnswering(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            start_positions=sequence_labels,
+            end_positions=sequence_labels,
+        )
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+
+    def create_and_check_for_sequence_classification(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        config.num_labels = self.num_labels
+        model = RoCBertForSequenceClassification(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            labels=sequence_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def create_and_check_for_token_classification(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        config.num_labels = self.num_labels
+        model = RoCBertForTokenClassification(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            input_shape_ids=input_shape_ids,
+            input_pronunciation_ids=input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            labels=token_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
+
+    def create_and_check_for_multiple_choice(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        config.num_choices = self.num_choices
+        model = RoCBertForMultipleChoice(config=config)
+        model.to(torch_device)
+        model.eval()
+        multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_inputs_shape_ids = input_shape_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_inputs_pronunciation_ids = (
+            input_pronunciation_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        )
+        multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        result = model(
+            multiple_choice_inputs_ids,
+            input_shape_ids=multiple_choice_inputs_shape_ids,
+            input_pronunciation_ids=multiple_choice_inputs_pronunciation_ids,
+            attention_mask=multiple_choice_input_mask,
+            token_type_ids=multiple_choice_token_type_ids,
+            labels=choice_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            input_shape_ids,
+            input_pronunciation_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "input_shape_ids": input_shape_ids,
+            "input_pronunciation_ids": input_pronunciation_ids,
+            "token_type_ids": token_type_ids,
+            "attention_mask": input_mask,
+        }
+        return config, inputs_dict
+
+    def create_and_check_for_pretraining(
+        self,
+        config,
+        input_ids,
+        input_shape_ids,
+        input_pronunciation_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+    ):
+        model = RoCBertForPreTraining(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            input_shape_ids,
+            input_pronunciation_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            attack_input_ids=input_ids,
+            attack_input_shape_ids=input_shape_ids,
+            attack_input_pronunciation_ids=input_pronunciation_ids,
+            attack_attention_mask=input_mask,
+            attack_token_type_ids=token_type_ids,
+            labels_input_ids=token_labels,
+            labels_input_shape_ids=input_shape_ids,
+            labels_input_pronunciation_ids=input_pronunciation_ids,
+            labels_attention_mask=input_mask,
+            labels_token_type_ids=token_type_ids,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+
+@require_torch
+class RoCBertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            RoCBertModel,
+            RoCBertForMaskedLM,
+            RoCBertForCausalLM,
+            RoCBertForMultipleChoice,
+            RoCBertForQuestionAnswering,
+            RoCBertForSequenceClassification,
+            RoCBertForTokenClassification,
+            RoCBertForPreTraining,
+        )
+        if is_torch_available()
+        else ()
+    )
+    # Doesn't run generation tests. There are interface mismatches when using `generate` -- TODO @gante
+    all_generative_model_classes = ()
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": RoCBertModel,
+            "fill-mask": RoCBertForMaskedLM,
+            "question-answering": RoCBertForQuestionAnswering,
+            "text-classification": RoCBertForSequenceClassification,
+            "text-generation": RoCBertForCausalLM,
+            "token-classification": RoCBertForTokenClassification,
+            "zero-shot": RoCBertForSequenceClassification,
+        }
+        if is_torch_available()
+        else {}
+    )
+
+    # TODO: Fix the failed tests when this model gets more usage
+    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 in [
+            "FillMaskPipelineTests",
+            "FeatureExtractionPipelineTests",
+            "TextClassificationPipelineTests",
+            "TokenClassificationPipelineTests",
+        ]:
+            # Get error: IndexError: index out of range in self.
+            # `word_shape_file` and `word_pronunciation_file` should be shrunk during tiny model creation,
+            # otherwise `IndexError` could occur in some embedding layers. Skip for now until this model has
+            # more usage.
+            return True
+
+        return False
+
+    # Overwriting to add `is_decoder` flag
+    def prepare_config_and_inputs_for_generate(self, batch_size=2):
+        config, inputs = super().prepare_config_and_inputs_for_generate(batch_size)
+        config.is_decoder = True
+        return config, inputs
+
+    # special case for ForPreTraining model
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
+
+        if return_labels:
+            if model_class in get_values(MODEL_FOR_PRETRAINING_MAPPING):
+                inputs_dict["labels_input_ids"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
+                )
+                inputs_dict["labels_input_shape_ids"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
+                )
+                inputs_dict["labels_input_pronunciation_ids"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
+                )
+                inputs_dict["attack_input_ids"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
+                )
+                inputs_dict["attack_input_shape_ids"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
+                )
+                inputs_dict["attack_input_pronunciation_ids"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
+                )
+        return inputs_dict
+
+    def setUp(self):
+        self.model_tester = RoCBertModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=RoCBertConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_various_embeddings(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        for type in ["absolute", "relative_key", "relative_key_query"]:
+            config_and_inputs[0].position_embedding_type = type
+            config_and_inputs[0]._attn_implementation = "eager"
+            self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
+
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def test_decoder_model_past_with_large_inputs_relative_pos_emb(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        config_and_inputs[0].position_embedding_type = "relative_key"
+        config_and_inputs[0]._attn_implementation = "eager"
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
+
+    def test_for_pretraining(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_pretraining(*config_and_inputs)
+
+    def test_model_as_decoder(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)
+
+    def test_model_as_decoder_with_default_input_mask(self):
+        (
+            config,
+            input_ids,
+            input_shape_ids,
+            input_pronunciation_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        ) = self.model_tester.prepare_config_and_inputs_for_decoder()
+
+        input_mask = None
+
+        self.model_tester.create_and_check_model_as_decoder(
+            config,
+            input_ids,
+            input_shape_ids,
+            input_pronunciation_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+    @slow
+    def test_model_from_pretrained(self):
+        model_name = "weiweishi/roc-bert-base-zh"
+        model = RoCBertModel.from_pretrained(model_name)
+        self.assertIsNotNone(model)
+
+    def attention_mask_padding_matches_padding_free_with_position_ids(
+        self, attn_implementation: str, fa_kwargs: bool = False
+    ):
+        """
+        Overwritten to account for the embeddings that rely on position ids.
+        """
+        if not self.has_attentions:
+            self.skipTest(reason="Model architecture does not support attentions")
+
+        max_new_tokens = 30
+        support_flag = {
+            "sdpa": "_supports_sdpa",
+            "flash_attention_2": "_supports_flash_attn",
+            "flash_attention_3": "_supports_flash_attn",
+        }
+
+        for model_class in self.all_generative_model_classes:
+            if attn_implementation != "eager" and not getattr(model_class, support_flag[attn_implementation]):
+                self.skipTest(f"{model_class.__name__} does not support {attn_implementation}")
+
+            # can't infer if new attn mask API is supported by assume that only model with attention backend support it
+            if not model_class._supports_attention_backend:
+                self.skipTest(f"{model_class.__name__} does not support new attention mask API")
+
+            if model_class._is_stateful:  # non-transformer models most probably have no packing support
+                self.skipTest(f"{model_class.__name__} doesn't support packing!")
+
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            if config.is_encoder_decoder:
+                self.skipTest("Model is an encoder-decoder")
+
+            if 0 not in inputs_dict.get("attention_mask", []) or "attention_mask" not in inputs_dict:
+                self.skipTest("Model dummy inputs should contain padding in their attention mask")
+
+            if "input_ids" not in inputs_dict or inputs_dict["input_ids"].ndim != 2:
+                self.skipTest("Model dummy inputs should contain text input ids")
+
+            # make sure that all models have enough positions for generation
+            dummy_input_ids = inputs_dict["input_ids"]
+            if hasattr(config, "max_position_embeddings"):
+                config.max_position_embeddings = max_new_tokens + dummy_input_ids.shape[1] + 1
+
+            model = model_class(config)
+            if "position_ids" not in inspect.signature(model.forward).parameters:
+                self.skipTest("Model does not support position_ids")
+
+            if (not fa_kwargs) and "position_ids" not in inspect.signature(model.forward).parameters:
+                continue  # this model doesn't accept position ids as input
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+
+                # Drop all keys except for the minimal set. Hard to manipulate with multimodals/head_mask/etc
+                inputs_dict = {k: v for k, v in inputs_dict.items() if k in ["input_ids", "attention_mask"]}
+
+                # Ensure left padding, to adapt for some models
+                if 0 in inputs_dict["attention_mask"][:, -1]:
+                    inputs_dict["attention_mask"] = inputs_dict["attention_mask"].flip(1)
+                dummy_attention_mask = inputs_dict["attention_mask"]
+                dummy_input_ids[~dummy_attention_mask.bool()] = config.get_text_config().pad_token_id
+
+                # Main difference to other models, we need to prepare position ids according to the attention mask
+                # as we use it to extract embeddings that rely on the correct position - naively increasing sequences do
+                # not suffice anymore atp. The solution here calculates an increasing sequences for all 1s and puts 0s else.
+                inputs_dict["position_ids"] = ((inputs_dict["attention_mask"] == 1).long().cumsum(dim=1) - 1) * (
+                    inputs_dict["attention_mask"] == 1
+                ).long()
+
+                model = (
+                    model_class.from_pretrained(
+                        tmpdirname,
+                        dtype=torch.bfloat16,
+                        attn_implementation=attn_implementation,
+                    )
+                    .to(torch_device)
+                    .eval()
+                )
+
+                if fa_kwargs:
+                    # flatten
+                    features = [
+                        {"input_ids": i[a.bool()].tolist()} for i, a in zip(dummy_input_ids, dummy_attention_mask)
+                    ]
+
+                    # add position_ids + fa_kwargs
+                    data_collator = DataCollatorWithFlattening(return_tensors="pt", return_flash_attn_kwargs=True)
+                    batch = data_collator(features)
+                    padfree_inputs_dict = {
+                        k: t.to(torch_device) if torch.is_tensor(t) else t for k, t in batch.items()
+                    }
+                else:
+                    # create packed position_ids
+                    position_ids = (
+                        torch.cat([torch.arange(length) for length in dummy_attention_mask.sum(1).tolist()])
+                        .long()
+                        .unsqueeze(0)
+                        .to(torch_device)
+                    )
+                    padfree_inputs_dict = {
+                        "input_ids": dummy_input_ids[dummy_attention_mask.bool()].unsqueeze(0),
+                        "position_ids": position_ids,
+                    }
+
+                # We need to do simple forward without cache in order to trigger packed SDPA/flex/eager attention path
+                res_padded = model(**inputs_dict, use_cache=False)
+                res_padfree = model(**padfree_inputs_dict, use_cache=False)
+
+                logits_padded = res_padded.logits[dummy_attention_mask.bool()]
+                logits_padfree = res_padfree.logits[0]
+
+                # acceptable numerical instability
+                tol = torch.finfo(torch.bfloat16).eps
+                torch.testing.assert_close(logits_padded, logits_padfree, rtol=tol, atol=tol)
+
+    def flash_attn_inference_equivalence(
+        self, attn_implementation: str, padding_side: str, atol: float = 4e-2, rtol: float = 4e-2
+    ):
+        super().flash_attn_inference_equivalence(
+            attn_implementation,
+            padding_side,
+            # relaxing the tolerance here
+            atol=6e-2,
+            rtol=4e-2,
+        )
+
+
+@require_torch
+class RoCBertModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_masked_lm(self):
+        model = RoCBertForMaskedLM.from_pretrained("weiweishi/roc-bert-base-zh")
+
+        # input_text: ['[CLS]', 'b', 'a', '里', '系', '[MASK]', '国', '的', '首', '都', '[SEP]'] is the adversarial text
+        # of ['[CLS]', '巴', '黎', '是', '[MASK]', '国', '的', '首', '都', '[SEP]'], means
+        # "Paris is the [MASK] of France" in English
+        input_ids = torch.tensor([[101, 144, 143, 7027, 5143, 103, 1744, 4638, 7674, 6963, 102]])
+        input_shape_ids = torch.tensor([[2, 20324, 23690, 8740, 706, 1, 10900, 23343, 20205, 5850, 2]])
+        input_pronunciation_ids = torch.tensor([[2, 718, 397, 52, 61, 1, 168, 273, 180, 243, 2]])
+
+        output = model(input_ids, input_shape_ids, input_pronunciation_ids)
+        output_ids = torch.argmax(output.logits, dim=2)
+
+        # convert to tokens is: ['[CLS]', '巴', '*', '黎', '是', '法', '国', '的', '首', '都', '[SEP]']
+        expected_output = torch.tensor([[101, 2349, 115, 7944, 3221, 3791, 1744, 4638, 7674, 6963, 102]])
+
+        assert torch.allclose(output_ids, expected_output)

transformers/tests/models/roc_bert/test_tokenization_roc_bert.py

@@ -0,0 +1,321 @@
+# Copyright 2022 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 json
+import os
+import unittest
+
+from transformers.models.roc_bert.tokenization_roc_bert import (
+    VOCAB_FILES_NAMES,
+    RoCBertBasicTokenizer,
+    RoCBertTokenizer,
+    RoCBertWordpieceTokenizer,
+    _is_control,
+    _is_punctuation,
+    _is_whitespace,
+)
+from transformers.testing_utils import require_tokenizers, slow
+
+from ...test_tokenization_common import TokenizerTesterMixin, filter_non_english
+
+
+@require_tokenizers
+class BertTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+    from_pretrained_id = "weiweishi/roc-bert-base-zh"
+    tokenizer_class = RoCBertTokenizer
+    rust_tokenizer_class = None
+    test_rust_tokenizer = False
+    space_between_special_tokens = True
+    from_pretrained_filter = filter_non_english
+
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+
+        vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]", "你", "好", "是", "谁", "a", "b", "c", "d"]
+        word_shape = {}
+        word_pronunciation = {}
+        for i, value in enumerate(vocab_tokens):
+            word_shape[value] = i
+            word_pronunciation[value] = i
+        cls.vocab_file = os.path.join(cls.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        cls.word_shape_file = os.path.join(cls.tmpdirname, VOCAB_FILES_NAMES["word_shape_file"])
+        cls.word_pronunciation_file = os.path.join(cls.tmpdirname, VOCAB_FILES_NAMES["word_pronunciation_file"])
+        with open(cls.vocab_file, "w", encoding="utf-8") as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+        with open(cls.word_shape_file, "w", encoding="utf-8") as word_shape_writer:
+            json.dump(word_shape, word_shape_writer, ensure_ascii=False)
+        with open(cls.word_pronunciation_file, "w", encoding="utf-8") as word_pronunciation_writer:
+            json.dump(word_pronunciation, word_pronunciation_writer, ensure_ascii=False)
+
+    def test_full_tokenizer(self):
+        tokenizer = self.tokenizer_class(self.vocab_file, self.word_shape_file, self.word_pronunciation_file)
+
+        tokens = tokenizer.tokenize("你好[SEP]你是谁")
+        self.assertListEqual(tokens, ["你", "好", "[SEP]", "你", "是", "谁"])
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [5, 6, 2, 5, 7, 8])
+        self.assertListEqual(tokenizer.convert_tokens_to_shape_ids(tokens), [5, 6, 2, 5, 7, 8])
+        self.assertListEqual(tokenizer.convert_tokens_to_pronunciation_ids(tokens), [5, 6, 2, 5, 7, 8])
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_chinese with BasicTokenizer->RoCBertBasicTokenizer
+    def test_chinese(self):
+        tokenizer = RoCBertBasicTokenizer()
+
+        self.assertListEqual(tokenizer.tokenize("ah\u535a\u63a8zz"), ["ah", "\u535a", "\u63a8", "zz"])
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_basic_tokenizer_lower with BasicTokenizer->RoCBertBasicTokenizer
+    def test_basic_tokenizer_lower(self):
+        tokenizer = RoCBertBasicTokenizer(do_lower_case=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU?  "), ["hello", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00e9llo"), ["hello"])
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_basic_tokenizer_lower_strip_accents_false with BasicTokenizer->RoCBertBasicTokenizer
+    def test_basic_tokenizer_lower_strip_accents_false(self):
+        tokenizer = RoCBertBasicTokenizer(do_lower_case=True, strip_accents=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hällo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00e9llo"), ["h\u00e9llo"])
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_basic_tokenizer_lower_strip_accents_true with BasicTokenizer->RoCBertBasicTokenizer
+    def test_basic_tokenizer_lower_strip_accents_true(self):
+        tokenizer = RoCBertBasicTokenizer(do_lower_case=True, strip_accents=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hallo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00e9llo"), ["hello"])
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_basic_tokenizer_lower_strip_accents_default with BasicTokenizer->RoCBertBasicTokenizer
+    def test_basic_tokenizer_lower_strip_accents_default(self):
+        tokenizer = RoCBertBasicTokenizer(do_lower_case=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["hallo", "!", "how", "are", "you", "?"]
+        )
+        self.assertListEqual(tokenizer.tokenize("H\u00e9llo"), ["hello"])
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_basic_tokenizer_no_lower with BasicTokenizer->RoCBertBasicTokenizer
+    def test_basic_tokenizer_no_lower(self):
+        tokenizer = RoCBertBasicTokenizer(do_lower_case=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU?  "), ["HeLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_basic_tokenizer_no_lower_strip_accents_false with BasicTokenizer->RoCBertBasicTokenizer
+    def test_basic_tokenizer_no_lower_strip_accents_false(self):
+        tokenizer = RoCBertBasicTokenizer(do_lower_case=False, strip_accents=False)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["HäLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_basic_tokenizer_no_lower_strip_accents_true with BasicTokenizer->RoCBertBasicTokenizer
+    def test_basic_tokenizer_no_lower_strip_accents_true(self):
+        tokenizer = RoCBertBasicTokenizer(do_lower_case=False, strip_accents=True)
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHäLLo!how  \n Are yoU?  "), ["HaLLo", "!", "how", "Are", "yoU", "?"]
+        )
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_basic_tokenizer_respects_never_split_tokens with BasicTokenizer->RoCBertBasicTokenizer
+    def test_basic_tokenizer_respects_never_split_tokens(self):
+        tokenizer = RoCBertBasicTokenizer(do_lower_case=False, never_split=["[UNK]"])
+
+        self.assertListEqual(
+            tokenizer.tokenize(" \tHeLLo!how  \n Are yoU? [UNK]"), ["HeLLo", "!", "how", "Are", "yoU", "?", "[UNK]"]
+        )
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_wordpiece_tokenizer with WordpieceTokenizer->RoCBertWordpieceTokenizer
+    def test_wordpiece_tokenizer(self):
+        vocab_tokens = ["[UNK]", "[CLS]", "[SEP]", "want", "##want", "##ed", "wa", "un", "runn", "##ing"]
+
+        vocab = {}
+        for i, token in enumerate(vocab_tokens):
+            vocab[token] = i
+        tokenizer = RoCBertWordpieceTokenizer(vocab=vocab, unk_token="[UNK]")
+
+        self.assertListEqual(tokenizer.tokenize(""), [])
+
+        self.assertListEqual(tokenizer.tokenize("unwanted running"), ["un", "##want", "##ed", "runn", "##ing"])
+
+        self.assertListEqual(tokenizer.tokenize("unwantedX running"), ["[UNK]", "runn", "##ing"])
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_is_whitespace
+    def test_is_whitespace(self):
+        self.assertTrue(_is_whitespace(" "))
+        self.assertTrue(_is_whitespace("\t"))
+        self.assertTrue(_is_whitespace("\r"))
+        self.assertTrue(_is_whitespace("\n"))
+        self.assertTrue(_is_whitespace("\u00a0"))
+
+        self.assertFalse(_is_whitespace("A"))
+        self.assertFalse(_is_whitespace("-"))
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_is_control
+    def test_is_control(self):
+        self.assertTrue(_is_control("\u0005"))
+
+        self.assertFalse(_is_control("A"))
+        self.assertFalse(_is_control(" "))
+        self.assertFalse(_is_control("\t"))
+        self.assertFalse(_is_control("\r"))
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_is_punctuation
+    def test_is_punctuation(self):
+        self.assertTrue(_is_punctuation("-"))
+        self.assertTrue(_is_punctuation("$"))
+        self.assertTrue(_is_punctuation("`"))
+        self.assertTrue(_is_punctuation("."))
+
+        self.assertFalse(_is_punctuation("A"))
+        self.assertFalse(_is_punctuation(" "))
+
+    def test_clean_text(self):
+        tokenizer = self.get_tokenizer()
+
+        # Example taken from the issue https://github.com/huggingface/tokenizers/issues/340
+        self.assertListEqual([tokenizer.tokenize(t) for t in ["Test", "\xad", "test"]], [["[UNK]"], [], ["[UNK]"]])
+
+        if self.test_rust_tokenizer:
+            rust_tokenizer = self.get_rust_tokenizer()
+            self.assertListEqual(
+                [rust_tokenizer.tokenize(t) for t in ["Test", "\xad", "test"]], [["[UNK]"], [], ["[UNK]"]]
+            )
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_offsets_with_special_characters
+    def test_offsets_with_special_characters(self):
+        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)
+
+                sentence = f"A, naïve {tokenizer_r.mask_token} AllenNLP sentence."
+                tokens = tokenizer_r.encode_plus(
+                    sentence,
+                    return_attention_mask=False,
+                    return_token_type_ids=False,
+                    return_offsets_mapping=True,
+                    add_special_tokens=True,
+                )
+
+                do_lower_case = tokenizer_r.do_lower_case if hasattr(tokenizer_r, "do_lower_case") else False
+                expected_results = (
+                    [
+                        ((0, 0), tokenizer_r.cls_token),
+                        ((0, 1), "A"),
+                        ((1, 2), ","),
+                        ((3, 5), "na"),
+                        ((5, 6), "##ï"),
+                        ((6, 8), "##ve"),
+                        ((9, 15), tokenizer_r.mask_token),
+                        ((16, 21), "Allen"),
+                        ((21, 23), "##NL"),
+                        ((23, 24), "##P"),
+                        ((25, 33), "sentence"),
+                        ((33, 34), "."),
+                        ((0, 0), tokenizer_r.sep_token),
+                    ]
+                    if not do_lower_case
+                    else [
+                        ((0, 0), tokenizer_r.cls_token),
+                        ((0, 1), "a"),
+                        ((1, 2), ","),
+                        ((3, 8), "naive"),
+                        ((9, 15), tokenizer_r.mask_token),
+                        ((16, 21), "allen"),
+                        ((21, 23), "##nl"),
+                        ((23, 24), "##p"),
+                        ((25, 33), "sentence"),
+                        ((33, 34), "."),
+                        ((0, 0), tokenizer_r.sep_token),
+                    ]
+                )
+
+                self.assertEqual(
+                    [e[1] for e in expected_results], tokenizer_r.convert_ids_to_tokens(tokens["input_ids"])
+                )
+                self.assertEqual([e[0] for e in expected_results], tokens["offset_mapping"])
+
+    # Copied from tests.models.bert.test_tokenization_bert.BertTokenizationTest.test_change_tokenize_chinese_chars
+    def test_change_tokenize_chinese_chars(self):
+        list_of_common_chinese_char = ["的", "人", "有"]
+        text_with_chinese_char = "".join(list_of_common_chinese_char)
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                kwargs["tokenize_chinese_chars"] = True
+                tokenizer_p = self.get_tokenizer(pretrained_name, **kwargs)
+                tokenizer_r = self.get_rust_tokenizer(pretrained_name, **kwargs)
+
+                ids_without_spe_char_p = tokenizer_p.encode(text_with_chinese_char, add_special_tokens=False)
+                ids_without_spe_char_r = tokenizer_r.encode(text_with_chinese_char, add_special_tokens=False)
+
+                tokens_without_spe_char_r = tokenizer_r.convert_ids_to_tokens(ids_without_spe_char_r)
+                tokens_without_spe_char_p = tokenizer_p.convert_ids_to_tokens(ids_without_spe_char_p)
+
+                # it is expected that each Chinese character is not preceded by "##"
+                self.assertListEqual(tokens_without_spe_char_p, list_of_common_chinese_char)
+                self.assertListEqual(tokens_without_spe_char_r, list_of_common_chinese_char)
+
+                kwargs["tokenize_chinese_chars"] = False
+                tokenizer_r = self.get_rust_tokenizer(pretrained_name, **kwargs)
+                tokenizer_p = self.get_tokenizer(pretrained_name, **kwargs)
+
+                ids_without_spe_char_r = tokenizer_r.encode(text_with_chinese_char, add_special_tokens=False)
+                ids_without_spe_char_p = tokenizer_p.encode(text_with_chinese_char, add_special_tokens=False)
+
+                tokens_without_spe_char_r = tokenizer_r.convert_ids_to_tokens(ids_without_spe_char_r)
+                tokens_without_spe_char_p = tokenizer_p.convert_ids_to_tokens(ids_without_spe_char_p)
+
+                # it is expected that only the first Chinese character is not preceded by "##".
+                expected_tokens = [
+                    f"##{token}" if idx != 0 else token for idx, token in enumerate(list_of_common_chinese_char)
+                ]
+                self.assertListEqual(tokens_without_spe_char_p, expected_tokens)
+                self.assertListEqual(tokens_without_spe_char_r, expected_tokens)
+
+    @slow
+    def test_sequence_builders(self):
+        tokenizer = self.tokenizer_class(self.vocab_file, self.word_shape_file, self.word_pronunciation_file)
+
+        text = tokenizer.encode("你好", add_special_tokens=False)
+        text_2 = tokenizer.encode("你是谁", add_special_tokens=False)
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        assert encoded_sentence == [1] + text + [2]
+        assert encoded_pair == [1] + text + [2] + text_2 + [2]
+
+    def test_prepare_for_model(self):
+        tokenizers = self.get_tokenizers(do_lower_case=False)
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                string_sequence = "你好，你是谁"
+                tokens = tokenizer.tokenize(string_sequence)
+                tokens_ids = tokenizer.convert_tokens_to_ids(tokens)
+                tokens_shape_ids = tokenizer.convert_tokens_to_shape_ids(tokens)
+                tokens_proun_ids = tokenizer.convert_tokens_to_pronunciation_ids(tokens)
+                prepared_input_dict = tokenizer.prepare_for_model(
+                    tokens_ids, tokens_shape_ids, tokens_proun_ids, add_special_tokens=True
+                )
+
+                input_dict = tokenizer.encode_plus(string_sequence, add_special_tokens=True)
+
+                self.assertEqual(input_dict, prepared_input_dict)