first commit

tests/tokenization/test_tokenization_fast.py

@@ -0,0 +1,392 @@
+# Copyright 2019 HuggingFace Inc.
+#
+# 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 concurrent.futures
+import json
+import os
+import shutil
+import tempfile
+import unittest
+
+from tokenizers import Tokenizer, decoders, pre_tokenizers, trainers
+from tokenizers.models import BPE, WordLevel
+
+from transformers import AutoTokenizer, PreTrainedTokenizerFast
+from transformers.testing_utils import require_tokenizers
+
+
+@require_tokenizers
+class PreTrainedTokenizationFastTest(unittest.TestCase):
+    rust_tokenizer_class = PreTrainedTokenizerFast
+    from_pretrained_vocab_key = "tokenizer_file"
+
+    @classmethod
+    def setUpClass(cls):
+        cls.tmpdirname = tempfile.mkdtemp()
+        cls.model_paths = cls._create_test_tokenizers()
+        cls.bytelevel_bpe_model_name = cls.model_paths[1]
+        cls.tokenizers_list = [(cls.rust_tokenizer_class, path, {}) for path in cls.model_paths]
+
+    @classmethod
+    def tearDownClass(cls):
+        shutil.rmtree(cls.tmpdirname, ignore_errors=True)
+
+    @classmethod
+    def _create_test_tokenizers(cls):
+        paths = []
+
+        wordlevel_dir = os.path.join(cls.tmpdirname, "wordlevel_tokenizer")
+        os.makedirs(wordlevel_dir, exist_ok=True)
+        wl_vocab = {"[UNK]": 0, "[PAD]": 1, "hello": 2, "world": 3, "test": 4}
+        wordlevel_tokenizer = Tokenizer(WordLevel(wl_vocab, unk_token="[UNK]"))
+        wordlevel_tokenizer.pre_tokenizer = pre_tokenizers.Whitespace()
+        fast_wl = PreTrainedTokenizerFast(
+            tokenizer_object=wordlevel_tokenizer,
+            unk_token="[UNK]",
+            pad_token="[PAD]",
+            cls_token="[CLS]",
+            sep_token="[SEP]",
+            mask_token="[MASK]",
+        )
+        fast_wl.save_pretrained(wordlevel_dir)
+        paths.append(wordlevel_dir)
+
+        bpe_dir = os.path.join(cls.tmpdirname, "bytelevel_bpe_tokenizer")
+        os.makedirs(bpe_dir, exist_ok=True)
+        bpe_tokenizer = Tokenizer(BPE(unk_token="[UNK]"))
+        trainer = trainers.BpeTrainer(
+            special_tokens=["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"],
+            vocab_size=100,
+        )
+        corpus = ["Hello world!", "Test the byte level BPE tokenizer.", "Tokenizer fast test."]
+        bpe_tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel()
+        bpe_tokenizer.train_from_iterator(corpus, trainer=trainer)
+        bpe_tokenizer.decoder = decoders.ByteLevel()
+        fast_bpe = PreTrainedTokenizerFast(
+            tokenizer_object=bpe_tokenizer,
+            unk_token="[UNK]",
+            pad_token="[PAD]",
+            cls_token="[CLS]",
+            sep_token="[SEP]",
+            mask_token="[MASK]",
+        )
+        fast_bpe.save_pretrained(bpe_dir)
+        paths.append(bpe_dir)
+
+        return paths
+
+    @unittest.skip(
+        "We disable this test for PreTrainedTokenizerFast because it is the only tokenizer that is not linked to any model"
+    )
+    def test_encode_decode_with_spaces(self):
+        pass
+
+    @unittest.skip(
+        "We disable this test for PreTrainedTokenizerFast because it is the only tokenizer that is not linked to any model"
+    )
+    def test_added_tokens_serialization(self):
+        pass
+
+    @unittest.skip(
+        "We disable this test for PreTrainedTokenizerFast because it is the only tokenizer that is not linked to any model"
+    )
+    def test_additional_special_tokens_serialization(self):
+        pass
+
+    def test_training_new_tokenizer(self):
+        tmpdirname_orig = self.tmpdirname
+        # Here we want to test the 2 available tokenizers that use 2 different types of models: Unigram and WordLevel.
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                try:
+                    self.tmpdirname = tempfile.mkdtemp()
+                    tokenizer = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+
+                    tokenizer.save_pretrained(self.tmpdirname)
+                    reloaded = PreTrainedTokenizerFast.from_pretrained(self.tmpdirname)
+                    self.assertEqual(reloaded.get_vocab(), tokenizer.get_vocab())
+                finally:
+                    # Even if the test fails, we must be sure that the folder is deleted and that the default tokenizer
+                    # is restored
+                    shutil.rmtree(self.tmpdirname)
+                    self.tmpdirname = tmpdirname_orig
+
+    def test_training_new_tokenizer_with_special_tokens_change(self):
+        tmpdirname_orig = self.tmpdirname
+        # Here we want to test the 2 available tokenizers that use 2 different types of models: Unigram and WordLevel.
+        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
+            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
+                try:
+                    self.tmpdirname = tempfile.mkdtemp()
+                    tokenizer = self.rust_tokenizer_class.from_pretrained(pretrained_name, **kwargs)
+
+                    tokenizer.add_special_tokens({"pad_token": "<pad>"})
+                    tokenizer.save_pretrained(self.tmpdirname)
+                    reloaded = PreTrainedTokenizerFast.from_pretrained(self.tmpdirname)
+                    self.assertEqual(reloaded.pad_token, "<pad>")
+                finally:
+                    # Even if the test fails, we must be sure that the folder is deleted and that the default tokenizer
+                    # is restored
+                    shutil.rmtree(self.tmpdirname)
+                    self.tmpdirname = tmpdirname_orig
+
+    def test_training_new_tokenizer_with_bytelevel(self):
+        tokenizer = self.rust_tokenizer_class.from_pretrained(self.bytelevel_bpe_model_name)
+
+        toy_text_iterator = ("a" for _ in range(1000))
+        new_tokenizer = tokenizer.train_new_from_iterator(text_iterator=toy_text_iterator, length=1000, vocab_size=50)
+
+        encoding_ids = new_tokenizer.encode("a🤗")
+        self.assertGreater(len(encoding_ids), 0)
+        self.assertEqual(new_tokenizer.decode(encoding_ids), " a🤗")
+
+    def test_init_from_tokenizers_model(self):
+        from tokenizers import Tokenizer
+
+        sentences = ["Hello, y'all!", "How are you 😁 ? There should not be any issue right?"]
+
+        tokenizer = Tokenizer.from_pretrained("google-t5/t5-base")
+        # Enable padding
+        tokenizer.enable_padding(pad_id=0, pad_token="<pad>", length=512, pad_to_multiple_of=8)
+        self.assertEqual(
+            tokenizer.padding,
+            {
+                "length": 512,
+                "pad_to_multiple_of": 8,
+                "pad_id": 0,
+                "pad_token": "<pad>",
+                "pad_type_id": 0,
+                "direction": "right",
+            },
+        )
+        fast_tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer)
+        tmpdirname = tempfile.mkdtemp()
+        fast_tokenizer.save_pretrained(tmpdirname)
+        fast_from_saved = PreTrainedTokenizerFast.from_pretrained(tmpdirname)
+        for tok in [fast_tokenizer, fast_from_saved]:
+            self.assertEqual(tok.pad_token_id, 0)
+            self.assertEqual(tok.padding_side, "right")
+            self.assertEqual(tok.pad_token, "<pad>")
+            self.assertEqual(tok.init_kwargs["max_length"], 512)
+            self.assertEqual(tok.init_kwargs["pad_to_multiple_of"], 8)
+            self.assertEqual(tok(sentences, padding = True, return_token_type_ids=True), {'input_ids': [[8774, 6, 3, 63, 31, 1748, 55, 1, 0, 0, 0, 0,0, 0, 0, 0],[ 571, 33, 25, 3, 2, 3, 58, 290, 225, 59, 36, 136, 962, 269, 58, 1]], 'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0],[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]})  # fmt: skip
+
+        tokenizer.enable_truncation(8, stride=0, strategy="longest_first", direction="right")
+        self.assertEqual(
+            tokenizer.truncation, {"max_length": 8, "stride": 0, "strategy": "longest_first", "direction": "right"}
+        )
+        fast_tokenizer = PreTrainedTokenizerFast(tokenizer_object=tokenizer)
+        tmpdirname = tempfile.mkdtemp()
+        fast_tokenizer.save_pretrained(tmpdirname)
+        fast_from_saved = PreTrainedTokenizerFast.from_pretrained(tmpdirname)
+        for tok in [fast_tokenizer, fast_from_saved]:
+            self.assertEqual(tok.truncation_side, "right")
+            self.assertEqual(tok.init_kwargs["truncation_strategy"], "longest_first")
+            self.assertEqual(tok.init_kwargs["max_length"], 8)
+            self.assertEqual(tok.init_kwargs["stride"], 0)
+            # NOTE even if the model has a default max_length, it is not used...
+            # thus tok(sentences, truncation = True) does nothing and does not warn either
+            self.assertEqual(tok(sentences, truncation = True, max_length = 8, return_token_type_ids=True), {'input_ids': [[8774, 6, 3, 63, 31, 1748, 55, 1],[ 571, 33, 25, 3, 2, 3, 58, 1]], 'token_type_ids': [[0, 0, 0, 0, 0, 0, 0, 0],[0, 0, 0, 0, 0, 0, 0, 0]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1],[1, 1, 1, 1, 1, 1, 1, 1]]})  # fmt: skip
+
+    def test_class_after_save_and_reload(self):
+        model_id = self.model_paths[0]
+
+        with tempfile.TemporaryDirectory() as temp_dir:
+            tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True)
+            self.assertIsInstance(tokenizer, PreTrainedTokenizerFast)
+
+            tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
+            self.assertIsInstance(tokenizer, PreTrainedTokenizerFast)
+
+            tokenizer.save_pretrained(temp_dir)
+
+            tokenizer = AutoTokenizer.from_pretrained(temp_dir, use_fast=False)
+            self.assertIsInstance(tokenizer, PreTrainedTokenizerFast)
+
+            tokenizer = AutoTokenizer.from_pretrained(temp_dir, use_fast=True)
+            self.assertIsInstance(tokenizer, PreTrainedTokenizerFast)
+
+    def test_bpe_tokenizer_skips_clean_up_tokenization_spaces(self):
+        """BPE tokenizers should not apply clean_up_tokenization even when the flag is True.
+
+        clean_up_tokenization strips spaces before punctuation (e.g. " ." -> "."),
+        which was designed for WordPiece tokenizers. For BPE tokenizers, spaces are
+        encoded as part of tokens and the cleanup is destructive.
+        """
+        tokenizer = PreTrainedTokenizerFast.from_pretrained(self.bytelevel_bpe_model_name)
+        tokenizer.clean_up_tokenization_spaces = True
+
+        # Text with space before punctuation — cleanup would strip it if applied.
+        # Leading space accounts for ByteLevel BPE's add_prefix_space behavior.
+        text = " Hello world ."
+        ids = tokenizer.encode(text, add_special_tokens=False)
+        decoded = tokenizer.decode(ids, clean_up_tokenization_spaces=True)
+
+        # The space before "." must be preserved — BPE guard skips the cleanup
+        self.assertEqual(decoded, text)
+
+    def test_bpe_override_forces_cleanup(self):
+        """The escape hatch flag forces cleanup even for BPE tokenizers."""
+        tokenizer = PreTrainedTokenizerFast.from_pretrained(self.bytelevel_bpe_model_name)
+        tokenizer.clean_up_tokenization_spaces = True
+        tokenizer.clean_up_tokenization_spaces_for_bpe_even_though_it_will_corrupt_output = True
+
+        text = " Hello world ."
+        ids = tokenizer.encode(text, add_special_tokens=False)
+        decoded = tokenizer.decode(ids, clean_up_tokenization_spaces=True)
+
+        # With the override, cleanup IS applied — spaces before punctuation are stripped
+        self.assertEqual(decoded, " Hello world.")
+
+    def test_bpe_override_irrelevant_when_cleanup_false(self):
+        """Override flag has no effect when clean_up_tokenization_spaces is False."""
+        tokenizer = PreTrainedTokenizerFast.from_pretrained(self.bytelevel_bpe_model_name)
+        tokenizer.clean_up_tokenization_spaces = False
+        tokenizer.clean_up_tokenization_spaces_for_bpe_even_though_it_will_corrupt_output = True
+
+        # Leading space accounts for ByteLevel BPE's add_prefix_space behavior
+        text = " Hello world ."
+        ids = tokenizer.encode(text, add_special_tokens=False)
+        decoded = tokenizer.decode(ids)
+
+        # cleanup=False takes precedence — text is preserved, override is irrelevant
+        self.assertEqual(decoded, text)
+
+    def test_non_bpe_tokenizer_still_cleans_up(self):
+        """Non-BPE tokenizers should still apply cleanup normally."""
+        # model_paths[0] is a WordLevel tokenizer (non-BPE)
+        tokenizer = PreTrainedTokenizerFast.from_pretrained(self.model_paths[0])
+        tokenizer.clean_up_tokenization_spaces = True
+
+        text = "hello world ."
+        ids = tokenizer.encode(text, add_special_tokens=False)
+        decoded = tokenizer.decode(ids, clean_up_tokenization_spaces=True)
+
+        # Non-BPE: cleanup IS applied — space before "." is stripped
+        self.assertNotIn(" .", decoded)
+
+
+@require_tokenizers
+class TokenizerVersioningTest(unittest.TestCase):
+    def test_local_versioning(self):
+        tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-cased")
+        json_tokenizer = json.loads(tokenizer._tokenizer.to_str())
+        json_tokenizer["model"]["vocab"]["huggingface"] = len(tokenizer)
+
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            # Hack to save this in the tokenizer_config.json
+            tokenizer.init_kwargs["fast_tokenizer_files"] = ["tokenizer.4.0.0.json"]
+            tokenizer.save_pretrained(tmp_dir)
+            json.dump(json_tokenizer, open(os.path.join(tmp_dir, "tokenizer.4.0.0.json"), "w"))
+
+            # This should pick the new tokenizer file as the version of Transformers is > 4.0.0
+            new_tokenizer = AutoTokenizer.from_pretrained(tmp_dir)
+            self.assertEqual(len(new_tokenizer), len(tokenizer) + 1)
+            json_tokenizer = json.loads(new_tokenizer._tokenizer.to_str())
+            self.assertIn("huggingface", json_tokenizer["model"]["vocab"])
+
+            # Will need to be adjusted if we reach v42 and this test is still here.
+            # Should pick the old tokenizer file as the version of Transformers is < 4.0.0
+            shutil.move(os.path.join(tmp_dir, "tokenizer.4.0.0.json"), os.path.join(tmp_dir, "tokenizer.42.0.0.json"))
+            tokenizer.init_kwargs["fast_tokenizer_files"] = ["tokenizer.42.0.0.json"]
+            tokenizer.save_pretrained(tmp_dir)
+            new_tokenizer = AutoTokenizer.from_pretrained(tmp_dir)
+            self.assertEqual(len(new_tokenizer), len(tokenizer))
+            json_tokenizer = json.loads(new_tokenizer._tokenizer.to_str())
+            self.assertNotIn("huggingface", json_tokenizer["model"]["vocab"])
+
+    def test_repo_versioning(self):
+        # This repo has two tokenizer files, one for v4.0.0 and above with an added token, one for versions lower.
+        repo = "hf-internal-testing/test-two-tokenizers"
+
+        # This should pick the new tokenizer file as the version of Transformers is > 4.0.0
+        tokenizer = AutoTokenizer.from_pretrained(repo)
+        self.assertEqual(len(tokenizer), 28997)
+        json_tokenizer = json.loads(tokenizer._tokenizer.to_str())
+        self.assertIn("huggingface", json_tokenizer["model"]["vocab"])
+
+        # Testing an older version by monkey-patching the version in the module it's used.
+        from unittest.mock import patch
+
+        import transformers as old_transformers
+
+        # Matt: The old test modified the module level version numbers
+        # which was (I think) the cause of strange flaky tests depending on test ordering.
+        # Using a context manager ensures the version mutation doesn't leak out of this test
+        with patch.object(old_transformers.tokenization_utils_base, "__version__", "3.0.0"):
+            old_tokenizer = old_transformers.models.auto.AutoTokenizer.from_pretrained(repo)
+            self.assertEqual(len(old_tokenizer), 28996)
+            json_tokenizer = json.loads(old_tokenizer._tokenizer.to_str())
+            self.assertNotIn("huggingface", json_tokenizer["model"]["vocab"])
+
+
+@require_tokenizers
+class ReduceMutableBorrowTests(unittest.TestCase):
+    def test_async_share_tokenizer(self):
+        # See https://github.com/huggingface/transformers/pull/12550
+        # and https://github.com/huggingface/tokenizers/issues/537
+        tokenizer = PreTrainedTokenizerFast.from_pretrained("robot-test/dummy-tokenizer-wordlevel")
+        text = "The Matrix is a 1999 science fiction action film."
+
+        with concurrent.futures.ThreadPoolExecutor() as executor:
+            futures = [executor.submit(self.fetch, tokenizer, text) for i in range(10)]
+            return_value = [future.result() for future in futures]
+            self.assertEqual(return_value, [[1, 10, 0, 8, 0, 18, 0, 0, 0, 2] for i in range(10)])
+
+    def fetch(self, tokenizer, text):
+        return tokenizer.encode(text, truncation="longest_first", padding="longest")
+
+
+@require_tokenizers
+class PatchMistralRegexHubCallTests(unittest.TestCase):
+    """
+    Regression tests for https://github.com/huggingface/transformers/issues/44749 /
+    https://github.com/huggingface/transformers/issues/43502
+
+    `_patch_mistral_regex` used to unconditionally call `huggingface_hub.model_info`
+    for any tokenizer with vocab > 100k (e.g. Qwen3), causing a large slowdown
+    and breaking offline / local-path loading.
+    """
+
+    def _dummy_backend_tokenizer(self):
+        tok = Tokenizer(WordLevel({"[UNK]": 0, "a": 1}, unk_token="[UNK]"))
+        tok.pre_tokenizer = pre_tokenizers.Whitespace()
+        return tok
+
+    def test_local_files_only_skips_model_info(self):
+        from unittest.mock import patch
+
+        tok = self._dummy_backend_tokenizer()
+        with patch(
+            "huggingface_hub.model_info",
+            side_effect=AssertionError("model_info must not be called when local_files_only=True"),
+        ):
+            result = PreTrainedTokenizerFast._patch_mistral_regex(
+                tok,
+                "some/non-mistral-repo-id",
+                local_files_only=True,
+            )
+        self.assertIsNotNone(result)
+
+    def test_hub_error_does_not_break_init(self):
+        from unittest.mock import patch
+
+        tok = self._dummy_backend_tokenizer()
+        with patch("huggingface_hub.model_info", side_effect=RuntimeError("hub down")):
+            # Must not raise — a Hub failure should be swallowed for non-Mistral models.
+            result = PreTrainedTokenizerFast._patch_mistral_regex(
+                tok,
+                "not-a-real/hub-id",
+            )
+        self.assertIsNotNone(result)

tests/tokenization/test_tokenization_utils.py

@@ -0,0 +1,375 @@
+# Copyright 2018 HuggingFace Inc..
+#
+# 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.
+"""
+ruff: isort: skip_file
+"""
+
+import os
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers import (
+    AutoTokenizer,
+    BatchEncoding,
+    BertTokenizer,
+    LlamaTokenizer,
+    PreTrainedTokenizerFast,
+    PythonBackend,
+    TensorType,
+    TokenSpan,
+    is_tokenizers_available,
+)
+from transformers.models.gpt2.tokenization_gpt2 import GPT2Tokenizer
+from transformers.testing_utils import (
+    CaptureStderr,
+    require_sentencepiece,
+    require_tokenizers,
+    require_torch,
+    slow,
+)
+
+
+if is_tokenizers_available():
+    import tokenizers
+    from tokenizers import AddedToken, Tokenizer
+    from tokenizers.models import WordPiece
+
+
+class TokenizerUtilsTest(unittest.TestCase):
+    def check_tokenizer_from_pretrained(self, tokenizer_class):
+        # max_model_input_sizes is a legacy attribute that may not exist on all tokenizers
+        if not hasattr(tokenizer_class, "max_model_input_sizes"):
+            return
+
+        s3_models = list(tokenizer_class.max_model_input_sizes.keys())
+        for model_name in s3_models[:1]:
+            tokenizer = tokenizer_class.from_pretrained(model_name)
+            self.assertIsNotNone(tokenizer)
+            self.assertIsInstance(tokenizer, tokenizer_class)
+            self.assertIsInstance(tokenizer, PythonBackend)
+
+            for special_tok in tokenizer.all_special_tokens:
+                self.assertIsInstance(special_tok, str)
+                special_tok_id = tokenizer.convert_tokens_to_ids(special_tok)
+                self.assertIsInstance(special_tok_id, int)
+
+    @slow
+    def test_pretrained_tokenizers(self):
+        self.check_tokenizer_from_pretrained(GPT2Tokenizer)
+
+    def test_tensor_type_from_str(self):
+        self.assertEqual(TensorType("pt"), TensorType.PYTORCH)
+        self.assertEqual(TensorType("np"), TensorType.NUMPY)
+
+    @require_tokenizers
+    def test_batch_encoding_word_to_tokens(self):
+        tokenizer_r = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
+        encoded = tokenizer_r(["Test", "\xad", "test"], is_split_into_words=True)
+
+        self.assertEqual(encoded.word_to_tokens(0), TokenSpan(start=1, end=2))
+        self.assertEqual(encoded.word_to_tokens(1), None)
+        self.assertEqual(encoded.word_to_tokens(2), TokenSpan(start=2, end=3))
+
+    def test_batch_encoding_with_labels(self):
+        batch = BatchEncoding({"inputs": [[1, 2, 3], [4, 5, 6]], "labels": [0, 1]})
+        tensor_batch = batch.convert_to_tensors(tensor_type="np")
+        self.assertEqual(tensor_batch["inputs"].shape, (2, 3))
+        self.assertEqual(tensor_batch["labels"].shape, (2,))
+        # test converting the converted
+        with CaptureStderr() as cs:
+            tensor_batch = batch.convert_to_tensors(tensor_type="np")
+        self.assertFalse(len(cs.err), msg=f"should have no warning, but got {cs.err}")
+
+        batch = BatchEncoding({"inputs": [1, 2, 3], "labels": 0})
+        tensor_batch = batch.convert_to_tensors(tensor_type="np", prepend_batch_axis=True)
+        self.assertEqual(tensor_batch["inputs"].shape, (1, 3))
+        self.assertEqual(tensor_batch["labels"].shape, (1,))
+
+    @require_torch
+    def test_batch_encoding_with_labels_pt(self):
+        batch = BatchEncoding({"inputs": [[1, 2, 3], [4, 5, 6]], "labels": [0, 1]})
+        tensor_batch = batch.convert_to_tensors(tensor_type="pt")
+        self.assertEqual(tensor_batch["inputs"].shape, (2, 3))
+        self.assertEqual(tensor_batch["labels"].shape, (2,))
+        # test converting the converted
+        with CaptureStderr() as cs:
+            tensor_batch = batch.convert_to_tensors(tensor_type="pt")
+        self.assertFalse(len(cs.err), msg=f"should have no warning, but got {cs.err}")
+
+        batch = BatchEncoding({"inputs": [1, 2, 3], "labels": 0})
+        tensor_batch = batch.convert_to_tensors(tensor_type="pt", prepend_batch_axis=True)
+        self.assertEqual(tensor_batch["inputs"].shape, (1, 3))
+        self.assertEqual(tensor_batch["labels"].shape, (1,))
+
+    def test_padding_accepts_tensors(self):
+        features = [{"input_ids": np.array([0, 1, 2])}, {"input_ids": np.array([0, 1, 2, 3])}]
+        tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
+
+        batch = tokenizer.pad(features, padding=True)
+        self.assertTrue(isinstance(batch["input_ids"], np.ndarray))
+        self.assertEqual(batch["input_ids"].tolist(), [[0, 1, 2, tokenizer.pad_token_id], [0, 1, 2, 3]])
+        batch = tokenizer.pad(features, padding=True, return_tensors="np")
+        self.assertTrue(isinstance(batch["input_ids"], np.ndarray))
+        self.assertEqual(batch["input_ids"].tolist(), [[0, 1, 2, tokenizer.pad_token_id], [0, 1, 2, 3]])
+
+    @require_tokenizers
+    def test_decoding_single_token(self):
+        for tokenizer_class in [BertTokenizer, BertTokenizer]:
+            with self.subTest(f"{tokenizer_class}"):
+                tokenizer = tokenizer_class.from_pretrained("google-bert/bert-base-cased")
+
+                token_id = 2300
+                decoded_flat = tokenizer.decode(token_id)
+                decoded_list = tokenizer.decode([token_id])
+
+                self.assertEqual(decoded_flat, "Force")
+                self.assertEqual(decoded_list, "Force")
+
+                token_id = 0
+                decoded_flat = tokenizer.decode(token_id)
+                decoded_list = tokenizer.decode([token_id])
+
+                self.assertEqual(decoded_flat, "[PAD]")
+                self.assertEqual(decoded_list, "[PAD]")
+
+                last_item_id = tokenizer.vocab_size - 1
+                decoded_flat = tokenizer.decode(last_item_id)
+                decoded_list = tokenizer.decode([last_item_id])
+
+                self.assertEqual(decoded_flat, "##：")
+                self.assertEqual(decoded_list, "##：")
+
+    def test_extra_special_tokens_multimodal(self):
+        attribute_special_tokens_list = [
+            "bos_token",
+            "eos_token",
+            "unk_token",
+            "sep_token",
+            "pad_token",
+            "cls_token",
+            "mask_token",
+        ]
+        llama_tokenizer = LlamaTokenizer.from_pretrained("huggyllama/llama-7b")
+        llama_tokenizer._set_model_specific_special_tokens(
+            {
+                "boi_token": "<image_start>",
+                "eoi_token": "<image_end>",
+                "image_token": "<image>",
+            }
+        )
+        multimodal_special_tokens_list = attribute_special_tokens_list + ["boi_token", "eoi_token", "image_token"]
+        self.assertListEqual(llama_tokenizer.SPECIAL_TOKENS_ATTRIBUTES, multimodal_special_tokens_list)
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            llama_tokenizer.save_pretrained(tmpdirname)
+
+            # load back and check we have extra special tokens set
+            loaded_tokenizer = LlamaTokenizer.from_pretrained(tmpdirname)
+            multimodal_special_tokens_list = attribute_special_tokens_list + ["boi_token", "eoi_token", "image_token"]
+            self.assertListEqual(loaded_tokenizer.SPECIAL_TOKENS_ATTRIBUTES, multimodal_special_tokens_list)
+
+            # We set an image_token_id before, so we can get an "image_token" as str that matches the id
+            self.assertTrue(loaded_tokenizer.image_token == "<image>")
+            self.assertTrue(loaded_tokenizer.image_token_id == loaded_tokenizer.convert_tokens_to_ids("<image>"))
+
+        # save one more time and make sure the image token can get loaded back
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            loaded_tokenizer.save_pretrained(tmpdirname)
+            loaded_tokenizer_with_extra_tokens = LlamaTokenizer.from_pretrained(tmpdirname)
+            self.assertTrue(loaded_tokenizer_with_extra_tokens.image_token == "<image>")
+
+        # test that we can also indicate extra tokens during load time
+        extra_special_tokens = {
+            "boi_token": "<image_start>",
+            "eoi_token": "<image_end>",
+            "image_token": "<image>",
+        }
+        tokenizer = LlamaTokenizer.from_pretrained("huggyllama/llama-7b", extra_special_tokens=extra_special_tokens)
+        self.assertTrue(tokenizer.image_token == "<image>")
+        self.assertTrue(tokenizer.image_token_id == loaded_tokenizer.convert_tokens_to_ids("<image>"))
+
+    @require_tokenizers
+    def test_decoding_skip_special_tokens(self):
+        for tokenizer_class in [BertTokenizer, BertTokenizer]:
+            with self.subTest(f"{tokenizer_class}"):
+                tokenizer = tokenizer_class.from_pretrained("google-bert/bert-base-cased")
+                tokenizer.add_tokens(["ஐ"], special_tokens=True)
+
+                # test special token with other tokens, skip the special tokens
+                sentence = "This is a beautiful flower ஐ"
+                ids = tokenizer(sentence)["input_ids"]
+                decoded_sent = tokenizer.decode(ids, skip_special_tokens=True)
+                self.assertEqual(decoded_sent, "This is a beautiful flower")
+
+                # test special token with other tokens, do not skip the special tokens
+                ids = tokenizer(sentence)["input_ids"]
+                decoded_sent = tokenizer.decode(ids, skip_special_tokens=False)
+                self.assertEqual(decoded_sent, "[CLS] This is a beautiful flower ஐ [SEP]")
+
+                # test special token stand alone, skip the special tokens
+                sentence = "ஐ"
+                ids = tokenizer(sentence)["input_ids"]
+                decoded_sent = tokenizer.decode(ids, skip_special_tokens=True)
+                self.assertEqual(decoded_sent, "")
+
+                # test special token stand alone, do not skip the special tokens
+                ids = tokenizer(sentence)["input_ids"]
+                decoded_sent = tokenizer.decode(ids, skip_special_tokens=False)
+                self.assertEqual(decoded_sent, "[CLS] ஐ [SEP]")
+
+                # test single special token alone, skip
+                pad_id = 0
+                decoded_sent = tokenizer.decode(pad_id, skip_special_tokens=True)
+                self.assertEqual(decoded_sent, "")
+
+                # test single special token alone, do not skip
+                decoded_sent = tokenizer.decode(pad_id, skip_special_tokens=False)
+                self.assertEqual(decoded_sent, "[PAD]")
+
+    @require_torch
+    def test_padding_accepts_tensors_pt(self):
+        import torch
+
+        features = [{"input_ids": torch.tensor([0, 1, 2])}, {"input_ids": torch.tensor([0, 1, 2, 3])}]
+        tokenizer = BertTokenizer.from_pretrained("google-bert/bert-base-cased")
+
+        batch = tokenizer.pad(features, padding=True)
+        self.assertTrue(isinstance(batch["input_ids"], torch.Tensor))
+        self.assertEqual(batch["input_ids"].tolist(), [[0, 1, 2, tokenizer.pad_token_id], [0, 1, 2, 3]])
+        batch = tokenizer.pad(features, padding=True, return_tensors="pt")
+        self.assertTrue(isinstance(batch["input_ids"], torch.Tensor))
+        self.assertEqual(batch["input_ids"].tolist(), [[0, 1, 2, tokenizer.pad_token_id], [0, 1, 2, 3]])
+
+    @require_tokenizers
+    def test_instantiation_from_tokenizers(self):
+        bert_tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
+        PreTrainedTokenizerFast(tokenizer_object=bert_tokenizer)
+
+    @require_tokenizers
+    def test_instantiation_from_tokenizers_json_file(self):
+        bert_tokenizer = Tokenizer(WordPiece(unk_token="[UNK]"))
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            bert_tokenizer.save(os.path.join(tmpdirname, "tokenizer.json"))
+            PreTrainedTokenizerFast(tokenizer_file=os.path.join(tmpdirname, "tokenizer.json"))
+
+    def test_len_tokenizer(self):
+        for tokenizer_class in [BertTokenizer, BertTokenizer]:
+            with self.subTest(f"{tokenizer_class}"):
+                tokenizer = tokenizer_class.from_pretrained("bert-base-uncased")
+                added_tokens_size = len(tokenizer.added_tokens_decoder)
+                self.assertEqual(len(tokenizer), tokenizer.vocab_size)
+
+                tokenizer.add_tokens(["<test_token>"])
+                self.assertEqual(len(tokenizer), tokenizer.vocab_size + 1)
+                self.assertEqual(len(tokenizer.added_tokens_decoder), added_tokens_size + 1)
+                self.assertEqual(len(tokenizer.added_tokens_encoder), added_tokens_size + 1)
+
+    @require_sentencepiece
+    def test_sentencepiece_cohabitation(self):
+        from sentencepiece import sentencepiece_model_pb2 as _original_protobuf  # noqa: F401
+
+        from transformers.convert_slow_tokenizer import import_protobuf  # noqa: F401
+
+        # Now this will try to import sentencepiece_model_pb2_new.py. This should not fail even if the protobuf
+        # was already imported.
+        import_protobuf()
+
+    def test_training_new_tokenizer_edge_cases(self):
+        _tokenizer = Tokenizer(tokenizers.models.BPE(vocab={"a": 1, "b": 2, "ab": 3}, merges=[("a", "b")]))
+        _tokenizer.pre_tokenizer = None
+
+        tokenizer = PreTrainedTokenizerFast(tokenizer_object=_tokenizer)
+        toy_text_iterator = ("a" for _ in range(1000))
+        tokenizer.train_new_from_iterator(text_iterator=toy_text_iterator, length=1000, vocab_size=50)
+
+        _tokenizer.normalizer = None
+        tokenizer = PreTrainedTokenizerFast(tokenizer_object=_tokenizer)
+        toy_text_iterator = ("a" for _ in range(1000))
+        tokenizer.train_new_from_iterator(text_iterator=toy_text_iterator, length=1000, vocab_size=50)
+
+        _tokenizer.post_processor = None
+        tokenizer = PreTrainedTokenizerFast(tokenizer_object=_tokenizer)
+        toy_text_iterator = ("a" for _ in range(1000))
+        tokenizer.train_new_from_iterator(text_iterator=toy_text_iterator, length=1000, vocab_size=50)
+
+    def test_encode_message(self):
+        tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
+        conversation = [
+            {"role": "system", "content": "You are a helpful assistant"},
+            {"role": "user", "content": "Hey there, how are you?"},
+            {"role": "assistant", "content": "Thank you for asking, I am doing well"},
+            {"role": "user", "content": "What's the weather like today?"},
+            {"role": "assistant", "content": "Today the weather is nice"},
+        ]
+
+        # First, test the default case, where we encode the whole conversation at once
+        whole_conversation_tokens = tokenizer.apply_chat_template(conversation, tokenize=True, return_dict=False)
+
+        # Now, test the message-by-message encoding
+        tokens = []
+        for i, message in enumerate(conversation):
+            tokens += tokenizer.encode_message_with_chat_template(message, conversation_history=conversation[:i])
+
+        self.assertEqual(whole_conversation_tokens, tokens)
+
+    def test_encode_message_raises_on_add_generation_prompt(self):
+        tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
+        conversation = [
+            {"role": "system", "content": "You are a helpful assistant"},
+            {"role": "user", "content": "Hey there, how are you?"},
+        ]
+        with self.assertRaises(ValueError):
+            tokenizer.encode_message_with_chat_template(conversation[0], add_generation_prompt=True)
+
+    @require_tokenizers
+    def test_special_tokens_overwrite(self):
+        text_with_nonspecial_tokens = "there are 2 cats"  # '2' is originally special
+
+        tokenizer = LlamaTokenizer.from_pretrained("hf-internal-testing/Ernie4_5_Tokenizer")
+        # Overwrite special tokens 0-9 to non-special
+        tokenizer.add_tokens([AddedToken(f"{i}", normalized=False, special=False) for i in range(10)])
+        self.assertTrue(
+            tokenizer.decode(tokenizer.encode(text_with_nonspecial_tokens), skip_special_tokens=True)
+            == text_with_nonspecial_tokens
+        )
+
+        # Checking if this carries over even after saving and relaoding
+        tokenizer.save_pretrained("/tmp/ernie_tokenizer")
+        new_tokenizer = AutoTokenizer.from_pretrained("/tmp/ernie_tokenizer")
+        self.assertTrue(
+            new_tokenizer.decode(new_tokenizer.encode(text_with_nonspecial_tokens), skip_special_tokens=True)
+            == text_with_nonspecial_tokens
+        )
+
+    def test_import_protobuf_decode_error_without_protobuf(self):
+        from unittest.mock import patch
+
+        from transformers.tokenization_utils_base import import_protobuf_decode_error
+
+        with patch("transformers.tokenization_utils_base.is_protobuf_available", return_value=False):
+            result = import_protobuf_decode_error()
+            self.assertEqual(result, ())
+
+    def test_import_protobuf_decode_error_does_not_mask_exceptions(self):
+        from unittest.mock import patch
+
+        from transformers.tokenization_utils_base import import_protobuf_decode_error
+
+        with patch("transformers.tokenization_utils_base.is_protobuf_available", return_value=False):
+            with self.assertRaises(ValueError):
+                try:
+                    raise ValueError("real error")
+                except import_protobuf_decode_error():
+                    pass