init

transformers/tests/models/gpt2/test_modeling_gpt2.py

@@ -0,0 +1,578 @@
+# 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 unittest
+
+import pytest
+
+from transformers import GPT2Config, is_torch_available
+from transformers.testing_utils import (
+    Expectations,
+    cleanup,
+    require_flash_attn,
+    require_torch,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+from ...causal_lm_tester import CausalLMModelTest, CausalLMModelTester
+from ...test_modeling_common import floats_tensor, ids_tensor
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        GPT2DoubleHeadsModel,
+        GPT2ForQuestionAnswering,
+        GPT2ForSequenceClassification,
+        GPT2ForTokenClassification,
+        GPT2LMHeadModel,
+        GPT2Model,
+        GPT2Tokenizer,
+    )
+
+
+class GPT2ModelTester(CausalLMModelTester):
+    if is_torch_available():
+        config_class = GPT2Config
+        base_model_class = GPT2Model
+        causal_lm_class = GPT2LMHeadModel
+        sequence_classification_class = GPT2ForSequenceClassification
+        token_classification_class = GPT2ForTokenClassification
+        question_answering_class = GPT2ForQuestionAnswering
+
+    def __init__(
+        self,
+        parent,
+        use_token_type_ids=True,
+        num_choices=4,
+        **kwargs,
+    ):
+        super().__init__(parent, use_token_type_ids=use_token_type_ids, **kwargs)
+        self.num_choices = num_choices
+
+    def prepare_config_and_inputs(
+        self, extra_inputs=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
+    ):
+        # Overwritten: `GPT2DoubleHeadsModel` uses extra inputs
+        (config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels) = (
+            super().prepare_config_and_inputs()
+        )
+
+        if extra_inputs:
+            mc_token_ids = ids_tensor([self.batch_size, self.num_choices], self.seq_length)
+            head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
+            config_and_inputs = (
+                config,
+                input_ids,
+                input_mask,
+                head_mask,
+                token_type_ids,
+                mc_token_ids,
+                sequence_labels,
+                token_labels,
+                choice_labels,
+            )
+        else:
+            config_and_inputs = (
+                config,
+                input_ids,
+                token_type_ids,
+                input_mask,
+                sequence_labels,
+                token_labels,
+                choice_labels,
+            )
+
+        config = self.get_config(
+            scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
+            reorder_and_upcast_attn=reorder_and_upcast_attn,
+        )
+
+        return config_and_inputs
+
+    def get_config(self, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False):
+        # Overwritten: `GPT2Config` has extra flags and we want to test them
+        config = super().get_config()
+        config.scale_attn_by_inverse_layer_idx = scale_attn_by_inverse_layer_idx
+        config.reorder_and_upcast_attn = reorder_and_upcast_attn
+        return config
+
+    def prepare_config_and_inputs_for_common(self):
+        # Overwritten: we want `token_type_ids` as part of the common inputs
+        config_and_inputs = self.prepare_config_and_inputs(extra_inputs=True)
+        config, input_ids, _, head_mask, token_type_ids, _, _, _, _ = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "head_mask": head_mask}
+        return config, inputs_dict
+
+    def prepare_config_and_inputs_for_decoder(self):
+        # Extra function: used in `encoder_decoder` tests
+        (
+            config,
+            input_ids,
+            input_mask,
+            head_mask,
+            token_type_ids,
+            _,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = self.prepare_config_and_inputs(extra_inputs=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_mask,
+            head_mask,
+            token_type_ids,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+
+@require_torch
+class GPT2ModelTest(CausalLMModelTest, unittest.TestCase):
+    all_model_classes = (
+        (
+            GPT2Model,
+            GPT2LMHeadModel,
+            GPT2DoubleHeadsModel,
+            GPT2ForQuestionAnswering,
+            GPT2ForSequenceClassification,
+            GPT2ForTokenClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+    pipeline_model_mapping = (
+        {
+            "feature-extraction": GPT2Model,
+            "question-answering": GPT2ForQuestionAnswering,
+            "text-classification": GPT2ForSequenceClassification,
+            "text-generation": GPT2LMHeadModel,
+            "token-classification": GPT2ForTokenClassification,
+            "zero-shot": GPT2ForSequenceClassification,
+        }
+        if is_torch_available()
+        else {}
+    )
+    all_parallelizable_model_classes = (GPT2LMHeadModel, GPT2DoubleHeadsModel) if is_torch_available() else ()
+    fx_compatible = False  # Broken by attention refactor cc @Cyrilvallez
+    test_missing_keys = False
+    test_model_parallel = True
+    model_tester_class = GPT2ModelTester
+
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        # Overwritten: special case for DoubleHeads model
+        inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
+
+        if return_labels:
+            if model_class.__name__ == "GPT2DoubleHeadsModel":
+                inputs_dict["labels"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.num_choices, self.model_tester.seq_length),
+                    dtype=torch.long,
+                    device=torch_device,
+                )
+                inputs_dict["input_ids"] = inputs_dict["labels"]
+                inputs_dict["token_type_ids"] = inputs_dict["labels"]
+                inputs_dict["mc_token_ids"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.num_choices),
+                    dtype=torch.long,
+                    device=torch_device,
+                )
+                inputs_dict["mc_labels"] = torch.zeros(
+                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
+                )
+        return inputs_dict
+
+    def test_gpt2_double_lm_head_model(self):
+        # extra test: model-specific class
+        config_and_inputs = self.model_tester.prepare_config_and_inputs(extra_inputs=True)
+        config, input_ids, input_mask, _, token_type_ids, mc_token_ids, _, _, _ = config_and_inputs
+        model = GPT2DoubleHeadsModel(config)
+        model.to(torch_device)
+        model.eval()
+
+        multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
+        multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
+        multiple_choice_token_type_ids = (
+            token_type_ids.unsqueeze(1).expand(-1, self.model_tester.num_choices, -1).contiguous()
+        )
+
+        inputs = {
+            "input_ids": multiple_choice_inputs_ids,
+            "mc_token_ids": mc_token_ids,
+            "attention_mask": multiple_choice_input_mask,
+            "token_type_ids": multiple_choice_token_type_ids,
+            "labels": multiple_choice_inputs_ids,
+        }
+
+        result = model(**inputs)
+        self.assertEqual(result.loss.shape, ())
+        self.assertEqual(
+            result.logits.shape,
+            (
+                self.model_tester.batch_size,
+                self.model_tester.num_choices,
+                self.model_tester.seq_length,
+                self.model_tester.vocab_size,
+            ),
+        )
+        self.assertEqual(result.mc_logits.shape, (self.model_tester.batch_size, self.model_tester.num_choices))
+
+    def test_gpt2_scale_attn_by_inverse_layer_idx(self):
+        # extra test: model-specific flag
+        config_and_inputs = self.model_tester.prepare_config_and_inputs(scale_attn_by_inverse_layer_idx=True)
+        config, input_ids, token_type_ids, _, _, _, _ = config_and_inputs
+
+        model = GPT2LMHeadModel(config)
+        model.to(torch_device)
+        result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
+        self.assertEqual(result.loss.shape, ())
+        self.assertEqual(
+            result.logits.shape,
+            (self.model_tester.batch_size, self.model_tester.seq_length, self.model_tester.vocab_size),
+        )
+        result.loss.backward()
+
+    def test_gpt2_reorder_and_upcast_attn(self):
+        # extra test: model-specific flag
+        config_and_inputs = self.model_tester.prepare_config_and_inputs(reorder_and_upcast_attn=True)
+        config, input_ids, token_type_ids, _, _, _, _ = config_and_inputs
+
+        model = GPT2LMHeadModel(config)
+        model.to(torch_device)
+        result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
+        self.assertEqual(result.loss.shape, ())
+        self.assertEqual(
+            result.logits.shape,
+            (self.model_tester.batch_size, self.model_tester.seq_length, self.model_tester.vocab_size),
+        )
+        result.loss.backward()
+
+    def test_training_gradient_checkpointing(self):
+        # overwritten: GPT2DoubleHeadsModel fails this test, non-standard class
+        self.original_all_model_classes = self.all_model_classes
+        self.all_model_classes = (cls for cls in self.all_model_classes if cls.__name__ != "GPT2DoubleHeadsModel")
+        super().test_training_gradient_checkpointing()
+        self.all_model_classes = self.original_all_model_classes
+
+    def test_training_gradient_checkpointing_use_reentrant(self):
+        # overwritten: GPT2DoubleHeadsModel fails this test, non-standard class
+        self.original_all_model_classes = self.all_model_classes
+        self.all_model_classes = (cls for cls in self.all_model_classes if cls.__name__ != "GPT2DoubleHeadsModel")
+        super().test_training_gradient_checkpointing_use_reentrant()
+        self.all_model_classes = self.original_all_model_classes
+
+    def test_training_gradient_checkpointing_use_reentrant_false(self):
+        # overwritten: GPT2DoubleHeadsModel fails this test, non-standard class
+        self.original_all_model_classes = self.all_model_classes
+        self.all_model_classes = (cls for cls in self.all_model_classes if cls.__name__ != "GPT2DoubleHeadsModel")
+        super().test_training_gradient_checkpointing_use_reentrant_false()
+        self.all_model_classes = self.original_all_model_classes
+
+
+@require_torch
+class GPT2ModelLanguageGenerationTest(unittest.TestCase):
+    def tearDown(self):
+        super().tearDown()
+        # clean-up as much as possible GPU memory occupied by PyTorch
+        cleanup(torch_device, gc_collect=True)
+
+    def _test_lm_generate_gpt2_helper(
+        self,
+        gradient_checkpointing=False,
+        reorder_and_upcast_attn=False,
+        scale_attn_by_inverse_layer_idx=False,
+        verify_outputs=True,
+    ):
+        model = GPT2LMHeadModel.from_pretrained(
+            "openai-community/gpt2",
+            reorder_and_upcast_attn=reorder_and_upcast_attn,
+            scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
+        )
+        if gradient_checkpointing:
+            model.gradient_checkpointing_enable()
+        else:
+            model.gradient_checkpointing_disable()
+        model.to(torch_device)
+
+        # The dog
+        input_ids = torch.tensor([[464, 3290]], dtype=torch.long, device=torch_device)
+
+        # The dog was found in a field near the intersection of West and West Streets.\n\nThe dog
+        expected_output_ids = [464, 3290, 373, 1043, 287, 257, 2214, 1474, 262, 16246, 286, 2688, 290, 2688, 27262, 13, 198, 198, 464, 3290,]  # fmt: skip
+        output_ids = model.generate(input_ids, do_sample=False, max_length=20)
+        if verify_outputs:
+            self.assertListEqual(output_ids[0].tolist(), expected_output_ids)
+
+    @slow
+    def test_lm_generate_gpt2(self):
+        self._test_lm_generate_gpt2_helper()
+
+    @slow
+    def test_lm_generate_gpt2_with_gradient_checkpointing(self):
+        self._test_lm_generate_gpt2_helper(gradient_checkpointing=True)
+
+    @slow
+    def test_lm_generate_gpt2_with_reorder_and_upcast_attn(self):
+        self._test_lm_generate_gpt2_helper(reorder_and_upcast_attn=True)
+
+    @slow
+    def test_lm_generate_gpt2_with_scale_attn_by_inverse_layer_idx(self):
+        self._test_lm_generate_gpt2_helper(scale_attn_by_inverse_layer_idx=True, verify_outputs=False)
+
+    @slow
+    def test_gpt2_sample(self):
+        tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2")
+        model = GPT2LMHeadModel.from_pretrained("openai-community/gpt2")
+        model.to(torch_device)
+
+        torch.manual_seed(0)
+        tokenized = tokenizer("Today is a nice day and", return_tensors="pt", return_token_type_ids=True)
+        input_ids = tokenized.input_ids.to(torch_device)
+        output_ids = model.generate(input_ids, do_sample=True, max_length=20)
+        output_str = tokenizer.decode(output_ids[0], skip_special_tokens=True)
+
+        token_type_ids = tokenized.token_type_ids.to(torch_device)
+        output_seq = model.generate(input_ids=input_ids, do_sample=True, num_return_sequences=5, max_length=20)
+        output_seq_tt = model.generate(
+            input_ids=input_ids, token_type_ids=token_type_ids, do_sample=True, num_return_sequences=5, max_length=20
+        )
+        output_seq_strs = tokenizer.batch_decode(output_seq, skip_special_tokens=True)
+        output_seq_tt_strs = tokenizer.batch_decode(output_seq_tt, skip_special_tokens=True)
+
+        expected_outputs = Expectations(
+            {
+                ("rocm", None): 'Today is a nice day and we can do this again."\n\nDana said that she will',
+                ("rocm", (9, 5)): "Today is a nice day and if you don't know anything about the state of play during your holiday",
+                ("cuda", None): "Today is a nice day and if you don't know anything about the state of play during your holiday",
+                ("xpu", 3): "Today is a nice day and if you don't know anything about the state of play during your holiday",
+            }
+        )  # fmt: skip
+        EXPECTED_OUTPUT = expected_outputs.get_expectation()
+        self.assertEqual(output_str, EXPECTED_OUTPUT)
+        self.assertTrue(
+            all(output_seq_strs[idx] != output_seq_tt_strs[idx] for idx in range(len(output_seq_tt_strs)))
+        )  # token_type_ids should change output
+
+    # TODO joao, manuel: remove this in v4.62.0
+    @slow
+    def test_contrastive_search_gpt2(self):
+        article = (
+            "DeepMind Technologies is a British artificial intelligence subsidiary of Alphabet Inc. and research "
+            "laboratory founded in 2010. DeepMind was acquired by Google in 2014. The company is based"
+        )
+
+        gpt2_tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2-large")
+        gpt2_model = GPT2LMHeadModel.from_pretrained("openai-community/gpt2-large").to(torch_device)
+        input_ids = gpt2_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
+
+        outputs = gpt2_model.generate(
+            input_ids,
+            penalty_alpha=0.6,
+            top_k=4,
+            max_length=256,
+            trust_remote_code=True,
+            custom_generate="transformers-community/contrastive-search",
+        )
+
+        generated_text = gpt2_tokenizer.batch_decode(outputs, skip_special_tokens=True)
+
+        self.assertListEqual(
+            generated_text,
+            [
+                "DeepMind Technologies is a British artificial intelligence subsidiary of Alphabet Inc. and research "
+                "laboratory founded in 2010. DeepMind was acquired by Google in 2014. The company is based in London, "
+                "United Kingdom\n\nGoogle has a lot of data on its users and uses it to improve its products, such as "
+                "Google Now, which helps users find the information they're looking for on the web. But the company "
+                "is not the only one to collect data on its users. Facebook, for example, has its own facial "
+                "recognition technology, as well as a database of millions of photos that it uses to personalize its "
+                "News Feed.\n\nFacebook's use of data is a hot topic in the tech industry, with privacy advocates "
+                "concerned about the company's ability to keep users' information private. In a blog post last "
+                'year, Facebook CEO Mark Zuckerberg said his company would "do our best to be transparent about our '
+                'data use and how we use it."\n\n"We have made it clear that we do not sell or share your data with '
+                'third parties," Zuckerberg wrote. "If you have questions or concerns, please reach out to us at '
+                'privacy@facebook.com."\n\nGoogle declined to comment on the privacy implications of its use of data, '
+                "but said in a statement to The Associated Press that"
+            ],
+        )
+
+    @require_flash_attn
+    @require_torch_gpu
+    @pytest.mark.flash_attn_test
+    @slow
+    def test_flash_attn_2_generate_padding_left(self):
+        """
+        Overwriting the common test as the test is flaky on tiny models
+        """
+        model = GPT2LMHeadModel.from_pretrained("gpt2", dtype=torch.float16).to(0)
+
+        tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+
+        texts = ["hi", "Hello this is a very long sentence"]
+
+        tokenizer.padding_side = "left"
+        tokenizer.pad_token = tokenizer.eos_token
+
+        inputs = tokenizer(texts, return_tensors="pt", padding=True).to(0)
+
+        output_native = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        output_native = tokenizer.batch_decode(output_native)
+
+        model = GPT2LMHeadModel.from_pretrained(
+            "gpt2", device_map={"": 0}, attn_implementation="flash_attention_2", dtype=torch.float16
+        )
+
+        output_fa_2 = model.generate(**inputs, max_new_tokens=20, do_sample=False)
+        output_fa_2 = tokenizer.batch_decode(output_fa_2)
+
+        expected_output = [
+            "<|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|><|endoftext|>hi, who was born in the city of Kolkata, was a member of the Kolkata",
+            "Hello this is a very long sentence. I'm sorry. I'm sorry. I'm sorry. I'm sorry. I'm sorry",
+        ]
+
+        self.assertListEqual(output_native, output_fa_2)
+        self.assertListEqual(output_native, expected_output)
+
+    @slow
+    def test_batch_generation(self):
+        model = GPT2LMHeadModel.from_pretrained("openai-community/gpt2")
+        model.to(torch_device)
+        tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2")
+
+        tokenizer.padding_side = "left"
+
+        # Define PAD Token = EOS Token = 50256
+        tokenizer.pad_token = tokenizer.eos_token
+        model.config.pad_token_id = model.config.eos_token_id
+
+        # use different length sentences to test batching
+        sentences = [
+            "Hello, my dog is a little",
+            "Today, I",
+        ]
+
+        inputs = tokenizer(sentences, return_tensors="pt", padding=True)
+        input_ids = inputs["input_ids"].to(torch_device)
+        token_type_ids = torch.cat(
+            [
+                input_ids.new_full((input_ids.shape[0], input_ids.shape[1] - 1), 0),
+                input_ids.new_full((input_ids.shape[0], 1), 500),
+            ],
+            dim=-1,
+        )
+
+        outputs = model.generate(
+            input_ids=input_ids,
+            attention_mask=inputs["attention_mask"].to(torch_device),
+            max_length=20,
+        )
+
+        outputs_tt = model.generate(
+            input_ids=input_ids,
+            attention_mask=inputs["attention_mask"].to(torch_device),
+            token_type_ids=token_type_ids,
+            max_length=20,
+        )
+
+        inputs_non_padded = tokenizer(sentences[0], return_tensors="pt").input_ids.to(torch_device)
+        output_non_padded = model.generate(input_ids=inputs_non_padded, max_length=20)
+
+        num_paddings = inputs_non_padded.shape[-1] - inputs["attention_mask"][-1].long().sum().item()
+        inputs_padded = tokenizer(sentences[1], return_tensors="pt").input_ids.to(torch_device)
+        output_padded = model.generate(input_ids=inputs_padded, max_length=model.config.max_length - num_paddings)
+
+        batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
+        batch_out_sentence_tt = tokenizer.batch_decode(outputs_tt, skip_special_tokens=True)
+        non_padded_sentence = tokenizer.decode(output_non_padded[0], skip_special_tokens=True)
+        padded_sentence = tokenizer.decode(output_padded[0], skip_special_tokens=True)
+
+        expected_output_sentence = [
+            "Hello, my dog is a little bit of a mess. I'm not sure if he's going",
+            "Today, I'm going to be doing a lot of research on this. I",
+        ]
+        self.assertListEqual(expected_output_sentence, batch_out_sentence)
+        self.assertTrue(batch_out_sentence_tt != batch_out_sentence)  # token_type_ids should change output
+        self.assertListEqual(expected_output_sentence, [non_padded_sentence, padded_sentence])
+
+    @slow
+    def test_batch_generation_2heads(self):
+        model = GPT2DoubleHeadsModel.from_pretrained("openai-community/gpt2")
+        model.to(torch_device)
+        tokenizer = GPT2Tokenizer.from_pretrained("openai-community/gpt2")
+
+        tokenizer.padding_side = "left"
+
+        # This tokenizer has no pad token, so we have to set it in some way
+        # Define PAD Token = EOS Token = 50256
+        tokenizer.pad_token = tokenizer.eos_token
+        model.config.pad_token_id = model.config.eos_token_id
+
+        # use different length sentences to test batching
+        sentences = [
+            "Hello, my dog is a little",
+            "Today, I",
+        ]
+
+        inputs = tokenizer(sentences, return_tensors="pt", padding=True)
+        input_ids = inputs["input_ids"].to(torch_device)
+        token_type_ids = torch.cat(
+            [
+                input_ids.new_full((input_ids.shape[0], input_ids.shape[1] - 1), 0),
+                input_ids.new_full((input_ids.shape[0], 1), 500),
+            ],
+            dim=-1,
+        )
+
+        outputs = model.generate(
+            input_ids=input_ids,
+            attention_mask=inputs["attention_mask"].to(torch_device),
+            max_length=20,
+        )
+
+        outputs_tt = model.generate(
+            input_ids=input_ids,
+            attention_mask=inputs["attention_mask"].to(torch_device),
+            token_type_ids=token_type_ids,
+            max_length=20,
+        )
+
+        inputs_non_padded = tokenizer(sentences[0], return_tensors="pt").input_ids.to(torch_device)
+        output_non_padded = model.generate(input_ids=inputs_non_padded, max_length=20)
+
+        num_paddings = inputs_non_padded.shape[-1] - inputs["attention_mask"][-1].long().sum().item()
+        inputs_padded = tokenizer(sentences[1], return_tensors="pt").input_ids.to(torch_device)
+        output_padded = model.generate(input_ids=inputs_padded, max_length=model.config.max_length - num_paddings)
+
+        batch_out_sentence = tokenizer.batch_decode(outputs, skip_special_tokens=True)
+        batch_out_sentence_tt = tokenizer.batch_decode(outputs_tt, skip_special_tokens=True)
+        non_padded_sentence = tokenizer.decode(output_non_padded[0], skip_special_tokens=True)
+        padded_sentence = tokenizer.decode(output_padded[0], skip_special_tokens=True)
+
+        expected_output_sentence = [
+            "Hello, my dog is a little bit of a mess. I'm not sure if he's going",
+            "Today, I'm going to be doing a lot of research on this. I",
+        ]
+        self.assertListEqual(expected_output_sentence, batch_out_sentence)
+        self.assertTrue(batch_out_sentence_tt != batch_out_sentence)  # token_type_ids should change output
+        self.assertListEqual(expected_output_sentence, [non_padded_sentence, padded_sentence])

transformers/tests/models/gpt2/test_tokenization_gpt2.py

@@ -0,0 +1,376 @@
+# 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 json
+import os
+import unittest
+
+from transformers import AutoTokenizer, GPT2Tokenizer, GPT2TokenizerFast
+from transformers.models.gpt2.tokenization_gpt2 import VOCAB_FILES_NAMES
+from transformers.testing_utils import require_jinja, require_tiktoken, require_tokenizers
+
+from ...test_tokenization_common import TokenizerTesterMixin
+
+
+@require_tokenizers
+class GPT2TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+    from_pretrained_id = "openai-community/gpt2"
+    tokenizer_class = GPT2Tokenizer
+    rust_tokenizer_class = GPT2TokenizerFast
+    test_rust_tokenizer = True
+    from_pretrained_kwargs = {"add_prefix_space": True}
+    test_seq2seq = False
+
+    @classmethod
+    def setUpClass(cls):
+        super().setUpClass()
+
+        # Adapted from Sennrich et al. 2015 and https://github.com/rsennrich/subword-nmt
+        vocab = [
+            "l",
+            "o",
+            "w",
+            "e",
+            "r",
+            "s",
+            "t",
+            "i",
+            "d",
+            "n",
+            "\u0120",
+            "\u0120l",
+            "\u0120n",
+            "\u0120lo",
+            "\u0120low",
+            "er",
+            "\u0120lowest",
+            "\u0120newer",
+            "\u0120wider",
+            "<unk>",
+            "<|endoftext|>",
+        ]
+        vocab_tokens = dict(zip(vocab, range(len(vocab))))
+        merges = ["#version: 0.2", "\u0120 l", "\u0120l o", "\u0120lo w", "e r", ""]
+        cls.special_tokens_map = {"unk_token": "<unk>"}
+
+        cls.vocab_file = os.path.join(cls.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        cls.merges_file = os.path.join(cls.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
+        with open(cls.vocab_file, "w", encoding="utf-8") as fp:
+            fp.write(json.dumps(vocab_tokens) + "\n")
+        with open(cls.merges_file, "w", encoding="utf-8") as fp:
+            fp.write("\n".join(merges))
+
+    @classmethod
+    def get_tokenizer(cls, pretrained_name=None, **kwargs):
+        kwargs.update(cls.special_tokens_map)
+        pretrained_name = pretrained_name or cls.tmpdirname
+        return GPT2Tokenizer.from_pretrained(pretrained_name, **kwargs)
+
+    @classmethod
+    def get_rust_tokenizer(cls, pretrained_name=None, **kwargs):
+        kwargs.update(cls.special_tokens_map)
+        pretrained_name = pretrained_name or cls.tmpdirname
+        return GPT2TokenizerFast.from_pretrained(pretrained_name, **kwargs)
+
+    def get_input_output_texts(self, tokenizer):
+        input_text = "lower newer"
+        output_text = "lower newer"
+        return input_text, output_text
+
+    def test_full_tokenizer(self):
+        tokenizer = GPT2Tokenizer(self.vocab_file, self.merges_file, **self.special_tokens_map)
+        text = "lower newer"
+        bpe_tokens = ["\u0120low", "er", "\u0120", "n", "e", "w", "er"]
+        tokens = tokenizer.tokenize(text, add_prefix_space=True)
+        self.assertListEqual(tokens, bpe_tokens)
+
+        input_tokens = tokens + [tokenizer.unk_token]
+        input_bpe_tokens = [14, 15, 10, 9, 3, 2, 15, 19]
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+
+    def test_rust_and_python_full_tokenizers(self):
+        if not self.test_rust_tokenizer:
+            self.skipTest(reason="test_rust_tokenizer is set to False")
+
+        tokenizer = self.get_tokenizer()
+        rust_tokenizer = self.get_rust_tokenizer(add_prefix_space=True)
+
+        sequence = "lower newer"
+
+        # Testing tokenization
+        tokens = tokenizer.tokenize(sequence, add_prefix_space=True)
+        rust_tokens = rust_tokenizer.tokenize(sequence)
+        self.assertListEqual(tokens, rust_tokens)
+
+        # Testing conversion to ids without special tokens
+        ids = tokenizer.encode(sequence, add_special_tokens=False, add_prefix_space=True)
+        rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
+        self.assertListEqual(ids, rust_ids)
+
+        # Testing conversion to ids with special tokens
+        rust_tokenizer = self.get_rust_tokenizer(add_prefix_space=True)
+        ids = tokenizer.encode(sequence, add_prefix_space=True)
+        rust_ids = rust_tokenizer.encode(sequence)
+        self.assertListEqual(ids, rust_ids)
+
+        # Testing the unknown token
+        input_tokens = tokens + [rust_tokenizer.unk_token]
+        input_bpe_tokens = [14, 15, 10, 9, 3, 2, 15, 19]
+        self.assertListEqual(rust_tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
+
+    @unittest.skip
+    def test_pretokenized_inputs(self, *args, **kwargs):
+        # It's very difficult to mix/test pretokenization with byte-level
+        # And get both GPT2 and Roberta to work at the same time (mostly an issue of adding a space before the string)
+        pass
+
+    def test_padding(self, max_length=15):
+        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)
+
+                # Simple input
+                s = "This is a simple input"
+                s2 = ["This is a simple input 1", "This is a simple input 2"]
+                p = ("This is a simple input", "This is a pair")
+                p2 = [
+                    ("This is a simple input 1", "This is a simple input 2"),
+                    ("This is a simple pair 1", "This is a simple pair 2"),
+                ]
+
+                # Simple input tests
+                self.assertRaises(ValueError, tokenizer_r.encode, s, max_length=max_length, padding="max_length")
+
+                # Simple input
+                self.assertRaises(ValueError, tokenizer_r.encode_plus, s, max_length=max_length, padding="max_length")
+
+                # Simple input
+                self.assertRaises(
+                    ValueError,
+                    tokenizer_r.batch_encode_plus,
+                    s2,
+                    max_length=max_length,
+                    padding="max_length",
+                )
+
+                # Pair input
+                self.assertRaises(ValueError, tokenizer_r.encode, p, max_length=max_length, padding="max_length")
+
+                # Pair input
+                self.assertRaises(ValueError, tokenizer_r.encode_plus, p, max_length=max_length, padding="max_length")
+
+                # Pair input
+                self.assertRaises(
+                    ValueError,
+                    tokenizer_r.batch_encode_plus,
+                    p2,
+                    max_length=max_length,
+                    padding="max_length",
+                )
+
+    def test_padding_if_pad_token_set_slow(self):
+        tokenizer = GPT2Tokenizer.from_pretrained(self.tmpdirname, pad_token="<pad>")
+
+        # Simple input
+        s = "This is a simple input"
+        s2 = ["This is a simple input looooooooong", "This is a simple input"]
+        p = ("This is a simple input", "This is a pair")
+        p2 = [
+            ("This is a simple input loooooong", "This is a simple input"),
+            ("This is a simple pair loooooong", "This is a simple pair"),
+        ]
+
+        pad_token_id = tokenizer.pad_token_id
+
+        out_s = tokenizer(s, padding="max_length", max_length=30, return_tensors="np")
+        out_s2 = tokenizer(s2, padding=True, truncate=True, return_tensors="np")
+        out_p = tokenizer(*p, padding="max_length", max_length=60, return_tensors="np")
+        out_p2 = tokenizer(p2, padding=True, truncate=True, return_tensors="np")
+
+        # s
+        # test single string max_length padding
+        self.assertEqual(out_s["input_ids"].shape[-1], 30)
+        self.assertTrue(pad_token_id in out_s["input_ids"])
+        self.assertTrue(0 in out_s["attention_mask"])
+
+        # s2
+        # test automatic padding
+        self.assertEqual(out_s2["input_ids"].shape[-1], 33)
+        # long slice doesn't have padding
+        self.assertFalse(pad_token_id in out_s2["input_ids"][0])
+        self.assertFalse(0 in out_s2["attention_mask"][0])
+        # short slice does have padding
+        self.assertTrue(pad_token_id in out_s2["input_ids"][1])
+        self.assertTrue(0 in out_s2["attention_mask"][1])
+
+        # p
+        # test single pair max_length padding
+        self.assertEqual(out_p["input_ids"].shape[-1], 60)
+        self.assertTrue(pad_token_id in out_p["input_ids"])
+        self.assertTrue(0 in out_p["attention_mask"])
+
+        # p2
+        # test automatic padding pair
+        self.assertEqual(out_p2["input_ids"].shape[-1], 52)
+        # long slice pair doesn't have padding
+        self.assertFalse(pad_token_id in out_p2["input_ids"][0])
+        self.assertFalse(0 in out_p2["attention_mask"][0])
+        # short slice pair does have padding
+        self.assertTrue(pad_token_id in out_p2["input_ids"][1])
+        self.assertTrue(0 in out_p2["attention_mask"][1])
+
+    def test_add_bos_token_slow(self):
+        bos_token = "$$$"
+        tokenizer = GPT2Tokenizer.from_pretrained(self.tmpdirname, bos_token=bos_token, add_bos_token=True)
+
+        s = "This is a simple input"
+        s2 = ["This is a simple input 1", "This is a simple input 2"]
+
+        bos_token_id = tokenizer.bos_token_id
+
+        out_s = tokenizer(s)
+        out_s2 = tokenizer(s2)
+
+        self.assertEqual(out_s.input_ids[0], bos_token_id)
+        self.assertTrue(all(o[0] == bos_token_id for o in out_s2.input_ids))
+
+        decode_s = tokenizer.decode(out_s.input_ids)
+        decode_s2 = tokenizer.batch_decode(out_s2.input_ids)
+
+        self.assertTrue(decode_s.startswith(bos_token))
+        self.assertTrue(all(d.startswith(bos_token) for d in decode_s2))
+
+    @unittest.skip(reason="tokenizer has no padding token")
+    def test_padding_different_model_input_name(self):
+        pass
+
+    def test_special_tokens_mask_input_pairs_and_bos_token(self):
+        # TODO: change to self.get_tokenizers() when the fast version is implemented
+        tokenizers = [self.get_tokenizer(do_lower_case=False, add_bos_token=True)]
+        for tokenizer in tokenizers:
+            with self.subTest(f"{tokenizer.__class__.__name__}"):
+                sequence_0 = "Encode this."
+                sequence_1 = "This one too please."
+                encoded_sequence = tokenizer.encode(sequence_0, add_special_tokens=False)
+                encoded_sequence += tokenizer.encode(sequence_1, add_special_tokens=False)
+                encoded_sequence_dict = tokenizer.encode_plus(
+                    sequence_0,
+                    sequence_1,
+                    add_special_tokens=True,
+                    return_special_tokens_mask=True,
+                )
+                encoded_sequence_w_special = encoded_sequence_dict["input_ids"]
+                special_tokens_mask = encoded_sequence_dict["special_tokens_mask"]
+                self.assertEqual(len(special_tokens_mask), len(encoded_sequence_w_special))
+
+                filtered_sequence = [
+                    (x if not special_tokens_mask[i] else None) for i, x in enumerate(encoded_sequence_w_special)
+                ]
+                filtered_sequence = [x for x in filtered_sequence if x is not None]
+                self.assertEqual(encoded_sequence, filtered_sequence)
+
+    @require_jinja
+    def test_tokenization_for_chat(self):
+        tokenizer = GPT2Tokenizer.from_pretrained(self.tmpdirname)
+        tokenizer.chat_template = "{% for message in messages %}{{ message.content }}{{ eos_token }}{% endfor %}"
+        test_chats = [
+            [{"role": "system", "content": "You are a helpful chatbot."}, {"role": "user", "content": "Hello!"}],
+            [
+                {"role": "system", "content": "You are a helpful chatbot."},
+                {"role": "user", "content": "Hello!"},
+                {"role": "assistant", "content": "Nice to meet you."},
+            ],
+            [{"role": "assistant", "content": "Nice to meet you."}, {"role": "user", "content": "Hello!"}],
+        ]
+        tokenized_chats = [tokenizer.apply_chat_template(test_chat) for test_chat in test_chats]
+        # fmt: off
+        expected_tokens = [[20, 1, 20, 10, 20, 4, 3, 10, 20, 10, 20, 3, 0, 20, 20, 20, 0, 10, 20, 20, 20, 6, 20, 1, 6, 20, 20, 20, 3, 0, 0, 1, 20, 20],
+                          [20, 1, 20, 10, 20, 4, 3, 10, 20, 10, 20, 3, 0, 20, 20, 20, 0, 10, 20, 20, 20, 6, 20, 1, 6, 20, 20, 20, 3, 0, 0, 1, 20, 20, 20, 7, 20, 3, 10, 6, 1, 10, 20, 3, 3, 6, 10, 20, 1, 20, 20, 20],
+                          [20, 7, 20, 3, 10, 6, 1, 10, 20, 3, 3, 6, 10, 20, 1, 20, 20, 20, 20, 3, 0, 0, 1, 20, 20]]
+        # fmt: on
+        for tokenized_chat, expected_tokens in zip(tokenized_chats, expected_tokens):
+            self.assertListEqual(tokenized_chat, expected_tokens)
+
+    @require_tiktoken
+    def test_tokenization_tiktoken(self):
+        from tiktoken import encoding_name_for_model
+
+        from transformers.integrations.tiktoken import convert_tiktoken_to_fast
+
+        encoding = encoding_name_for_model("gpt2")
+        convert_tiktoken_to_fast(encoding, self.tmpdirname)
+
+        tiktoken_fast_tokenizer = GPT2TokenizerFast.from_pretrained(self.tmpdirname)
+        rust_tokenizer = GPT2TokenizerFast.from_pretrained("openai-community/gpt2")
+        sequence = "lower newer"
+        self.assertEqual(
+            rust_tokenizer.decode(rust_tokenizer.encode(sequence)),
+            tiktoken_fast_tokenizer.decode(rust_tokenizer.encode(sequence)),
+        )
+
+
+@require_tokenizers
+class OPTTokenizationTest(unittest.TestCase):
+    def test_serialize_deserialize_fast_opt(self):
+        # More context:
+        # https://huggingface.co/wjmcat/opt-350m-paddle/discussions/1
+        # https://huggingface.slack.com/archives/C01N44FJDHT/p1653511495183519
+        # https://github.com/huggingface/transformers/pull/17088#discussion_r871246439
+
+        tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m", from_slow=True)
+        text = "A photo of a cat"
+
+        tokens_ids = tokenizer.encode(
+            text,
+        )
+        self.assertEqual(tokens_ids, [2, 250, 1345, 9, 10, 4758])
+        tokenizer.save_pretrained("test_opt")
+
+        tokenizer = AutoTokenizer.from_pretrained("./test_opt")
+        tokens_ids = tokenizer.encode(
+            text,
+        )
+        self.assertEqual(tokens_ids, [2, 250, 1345, 9, 10, 4758])
+
+    def test_fast_slow_equivalence(self):
+        tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m", use_slow=True)
+        text = "A photo of a cat"
+
+        tokens_ids = tokenizer.encode(
+            text,
+        )
+        # Same as above
+        self.assertEqual(tokens_ids, [2, 250, 1345, 9, 10, 4758])
+
+    @unittest.skip(reason="This test is failing because of a bug in the fast tokenizer")
+    def test_users_can_modify_bos(self):
+        tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m", from_slow=True)
+
+        tokenizer.bos_token = "bos"
+        tokenizer.bos_token_id = tokenizer.get_vocab()["bos"]
+
+        text = "A photo of a cat"
+        tokens_ids = tokenizer.encode(
+            text,
+        )
+        # We changed the bos token
+        self.assertEqual(tokens_ids, [31957, 250, 1345, 9, 10, 4758])
+        tokenizer.save_pretrained("./tok")
+        tokenizer = AutoTokenizer.from_pretrained("./tok")
+        self.assertTrue(tokenizer.is_fast)
+        tokens_ids = tokenizer.encode(
+            text,
+        )
+        self.assertEqual(tokens_ids, [31957, 250, 1345, 9, 10, 4758])