[Tokenizer] Fix slow and fast serialization (#26570)

.circleci/create_circleci_config.py

@@ -127,6 +127,7 @@ def to_dict(self):
             },
         ]
         steps.extend([{"run": l} for l in self.install_steps])
+        steps.extend([{"run": "pip install pytest-subtests"}])
         steps.append(
             {
                 "save_cache": {

src/transformers/convert_slow_tokenizer.py

@@ -1168,9 +1168,9 @@ def tokenizer(self, proto):
             )
             tokenizer.add_special_tokens(
                 [
-                    AddedToken("<unk>"),
-                    AddedToken("<s>"),
-                    AddedToken("</s>"),
+                    AddedToken("<unk>", normalized=False, special=True),
+                    AddedToken("<s>", normalized=False, special=True),
+                    AddedToken("</s>", normalized=False, special=True),
                 ]
             )
         else:

src/transformers/models/bart/tokenization_bart.py

@@ -204,8 +204,6 @@ def __init__(
         pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
 
         # Mask token behave like a normal word, i.e. include the space before it
-        # TODO seems like both slow and fast actually don't strip left and right soooooooo yeah. See `test_embeded_special_tokens`
-        # Also this not only will strip the spaces but any punctuation
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
         with open(vocab_file, encoding="utf-8") as vocab_handle:

src/transformers/models/bart/tokenization_bart_fast.py

@@ -170,7 +170,12 @@ def __init__(
         trim_offsets=True,
         **kwargs,
     ):
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+        # we have to specify that this tokens is special otherwise adding it will reset the normalized flag to `False` in `add_special_tokens`
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, normalized=True, special=True)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
         super().__init__(
             vocab_file,
             merges_file,

src/transformers/models/barthez/tokenization_barthez.py

@@ -138,8 +138,8 @@ def __init__(
         sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs,
     ) -> None:
-        # Mask token behave like a normal word, i.e. include the space before it
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+        # Mask token behave like a normal word, i.e. include the space before it. Will have normalized=False by default this way
+        mask_token = AddedToken(mask_token, lstrip=True, special=True) if isinstance(mask_token, str) else mask_token
 
         self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
 

src/transformers/models/bertweet/tokenization_bertweet.py

@@ -149,10 +149,10 @@ def __init__(
         self.merges_file = merges_file
 
         self.encoder = {}
-        self.encoder[bos_token] = 0
-        self.encoder[pad_token] = 1
-        self.encoder[eos_token] = 2
-        self.encoder[unk_token] = 3
+        self.encoder[str(bos_token)] = 0
+        self.encoder[str(pad_token)] = 1
+        self.encoder[str(eos_token)] = 2
+        self.encoder[str(unk_token)] = 3
 
         self.add_from_file(vocab_file)
 

src/transformers/models/camembert/tokenization_camembert.py

@@ -89,7 +89,7 @@ class CamembertTokenizer(PreTrainedTokenizer):
         mask_token (`str`, *optional*, defaults to `"<mask>"`):
             The token used for masking values. This is the token used when training this model with masked language
             modeling. This is the token which the model will try to predict.
-        additional_special_tokens (`List[str]`, *optional*, defaults to `["<s>NOTUSED", "</s>NOTUSED"]`):
+        additional_special_tokens (`List[str]`, *optional*, defaults to `['<s>NOTUSED', '</s>NOTUSED', '<unk>NOTUSED']`):
             Additional special tokens used by the tokenizer.
         sp_model_kwargs (`dict`, *optional*):
             Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
@@ -127,12 +127,16 @@ def __init__(
         unk_token="<unk>",
         pad_token="<pad>",
         mask_token="<mask>",
-        additional_special_tokens=["<s>NOTUSED", "</s>NOTUSED"],
+        additional_special_tokens=["<s>NOTUSED", "</s>NOTUSED", "<unk>NOTUSED"],
         sp_model_kwargs: Optional[Dict[str, Any]] = None,
         **kwargs,
     ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, rstrip=False, normalized=False, special=True)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
 
         self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
 
@@ -144,11 +148,11 @@ def __init__(
         # sentencepiece vocabulary (this is the case for <s> and </s> and <unk>).
         # In this case it is recommended to properly set the tokens by hand.
         self._added_tokens_decoder = {
-            0: AddedToken("<s>NOTUSED"),
-            1: AddedToken(pad_token),
-            2: AddedToken("</s>NOTUSED"),
-            3: AddedToken(unk_token),
-            4: AddedToken("<unk>NOTUSED"),
+            0: AddedToken("<s>NOTUSED", special=True),
+            1: AddedToken(pad_token, special=True) if isinstance(pad_token, str) else pad_token,
+            2: AddedToken("</s>NOTUSED", special=True),
+            3: AddedToken(unk_token, special=True) if isinstance(unk_token, str) else unk_token,
+            4: AddedToken("<unk>NOTUSED", special=True),
         }
 
         self.fairseq_offset = 4  # 3 tokens are newly added, but the offset starts from 4

src/transformers/models/camembert/tokenization_camembert_fast.py

@@ -119,12 +119,11 @@ def __init__(
         unk_token="<unk>",
         pad_token="<pad>",
         mask_token="<mask>",
-        additional_special_tokens=["<s>NOTUSED", "</s>NOTUSED"],
+        additional_special_tokens=["<s>NOTUSED", "</s>NOTUSED", "<unk>NOTUSED"],
         **kwargs,
     ):
-        # Mask token behave like a normal word, i.e. include the space before it
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
-
+        # Mask token behave like a normal word, i.e. include the space before it. Will have normalized = False
+        mask_token = AddedToken(mask_token, lstrip=True, special=True) if isinstance(mask_token, str) else mask_token
         super().__init__(
             vocab_file,
             tokenizer_file=tokenizer_file,

src/transformers/models/codegen/tokenization_codegen.py

@@ -163,10 +163,10 @@ def __init__(
         add_bos_token=False,
         **kwargs,
     ):
-        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
-        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
-        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
-        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        bos_token = AddedToken(bos_token, special=True) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, special=True) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, special=True) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, special=True) if isinstance(pad_token, str) else pad_token
         self.add_bos_token = add_bos_token
 
         with open(vocab_file, encoding="utf-8") as vocab_handle:

src/transformers/models/deberta/tokenization_deberta.py

@@ -192,12 +192,12 @@ def __init__(
         add_bos_token=False,
         **kwargs,
     ):
-        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
-        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
-        sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token
-        cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token
-        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
-        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        bos_token = AddedToken(bos_token, special=True) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, special=True) if isinstance(eos_token, str) else eos_token
+        sep_token = AddedToken(sep_token, special=True) if isinstance(sep_token, str) else sep_token
+        cls_token = AddedToken(cls_token, special=True) if isinstance(cls_token, str) else cls_token
+        unk_token = AddedToken(unk_token, special=True) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, special=True) if isinstance(pad_token, str) else pad_token
 
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token

src/transformers/models/deberta_v2/tokenization_deberta_v2.py

@@ -138,7 +138,7 @@ def __init__(
         self._tokenizer = SPMTokenizer(
             vocab_file, None, split_by_punct=split_by_punct, sp_model_kwargs=self.sp_model_kwargs
         )
-        unk_token = AddedToken(unk_token, normalized=True, lstrip=False, rstrip=False)
+        unk_token = AddedToken(unk_token, normalized=True, special=True) if isinstance(unk_token, str) else unk_token
         super().__init__(
             do_lower_case=do_lower_case,
             bos_token=bos_token,

src/transformers/models/fnet/tokenization_fnet.py

@@ -116,9 +116,10 @@ def __init__(
     ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it and
         # is included in the raw text, there should be a match in a non-normalized sentence.
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
-        cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token
-        sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token
+        mask_token = AddedToken(mask_token, lstrip=True, special=True) if isinstance(mask_token, str) else mask_token
+        cls_token = AddedToken(cls_token, special=True) if isinstance(cls_token, str) else cls_token
+        sep_token = AddedToken(sep_token, special=True) if isinstance(sep_token, str) else sep_token
+        mask_token = AddedToken(mask_token, special=True) if isinstance(mask_token, str) else mask_token
         self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
 
         self.do_lower_case = do_lower_case

src/transformers/models/layoutlmv2/tokenization_layoutlmv2.py

@@ -20,7 +20,7 @@
 import unicodedata
 from typing import Dict, List, Optional, Tuple, Union
 
-from ...tokenization_utils import PreTrainedTokenizer, _is_control, _is_punctuation, _is_whitespace
+from ...tokenization_utils import AddedToken, PreTrainedTokenizer, _is_control, _is_punctuation, _is_whitespace
 from ...tokenization_utils_base import (
     BatchEncoding,
     EncodedInput,
@@ -244,6 +244,12 @@ def __init__(
         additional_special_tokens: Optional[List[str]] = None,
         **kwargs,
     ):
+        sep_token = AddedToken(sep_token, special=True) if isinstance(sep_token, str) else sep_token
+        unk_token = AddedToken(unk_token, special=True) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, special=True) if isinstance(pad_token, str) else pad_token
+        cls_token = AddedToken(cls_token, special=True) if isinstance(cls_token, str) else cls_token
+        mask_token = AddedToken(mask_token, special=True) if isinstance(mask_token, str) else mask_token
+
         if not os.path.isfile(vocab_file):
             raise ValueError(
                 f"Can't find a vocabulary file at path '{vocab_file}'. To load the vocabulary from a Google pretrained"

src/transformers/models/layoutxlm/tokenization_layoutxlm.py

@@ -250,7 +250,7 @@ def __init__(
         **kwargs,
     ) -> None:
         # Mask token behave like a normal word, i.e. include the space before it
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+        mask_token = AddedToken(mask_token, lstrip=True, special=True) if isinstance(mask_token, str) else mask_token
 
         self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
 

src/transformers/models/led/tokenization_led.py

@@ -197,8 +197,6 @@ def __init__(
         pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
 
         # Mask token behave like a normal word, i.e. include the space before it
-        # TODO seems like both slow and fast actually don't strip left and right soooooooo yeah. See `test_embeded_special_tokens`
-        # Also this not only will strip the spaces but any punctuation
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
         with open(vocab_file, encoding="utf-8") as vocab_handle:

src/transformers/models/led/tokenization_led_fast.py

@@ -152,7 +152,12 @@ def __init__(
         trim_offsets=True,
         **kwargs,
     ):
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+        # we have to specify that this tokens is special otherwise adding it will reset the normalized flag to `False` in `add_special_tokens`
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, normalized=True, special=True)
+            if isinstance(mask_token, str)
+            else mask_token
+        )
         super().__init__(
             vocab_file,
             merges_file,

src/transformers/models/llama/tokenization_llama.py

@@ -118,10 +118,10 @@ def __init__(
         **kwargs,
     ):
         self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
-        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
-        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
-        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
-        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        bos_token = AddedToken(bos_token, normalized=False, special=True) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, normalized=False, special=True) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, normalized=False, special=True) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, normalized=False, special=True) if isinstance(pad_token, str) else pad_token
 
         if legacy is None:
             logger.warning_once(

src/transformers/models/marian/tokenization_marian.py

@@ -148,9 +148,9 @@ def __init__(
 
         self.separate_vocabs = separate_vocabs
         self.encoder = load_json(vocab)
-        if unk_token not in self.encoder:
+        if str(unk_token) not in self.encoder:
             raise KeyError("<unk> token must be in the vocab")
-        assert pad_token in self.encoder
+        assert str(pad_token) in self.encoder
 
         if separate_vocabs:
             self.target_encoder = load_json(target_vocab_file)

src/transformers/models/mbart/tokenization_mbart.py

@@ -97,7 +97,9 @@ def __init__(
         **kwargs,
     ):
         # Mask token behave like a normal word, i.e. include the space before it
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+        mask_token = (
+            AddedToken(mask_token, lstrip=True, normalized=False) if isinstance(mask_token, str) else mask_token
+        )
 
         self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
 

src/transformers/models/mbart50/tokenization_mbart50.py

@@ -132,7 +132,7 @@ def __init__(
 
         self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
 
-        kwargs["additional_special_tokens"] = kwargs.get("additional_special_tokens", [])
+        kwargs["additional_special_tokens"] = kwargs.get("additional_special_tokens", []) or []
         kwargs["additional_special_tokens"] += [
             code for code in FAIRSEQ_LANGUAGE_CODES if code not in kwargs["additional_special_tokens"]
         ]

src/transformers/models/mbart50/tokenization_mbart50_fast.py

@@ -127,7 +127,7 @@ def __init__(
         # Mask token behave like a normal word, i.e. include the space before it
         mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
 
-        kwargs["additional_special_tokens"] = kwargs.get("additional_special_tokens", [])
+        kwargs["additional_special_tokens"] = kwargs.get("additional_special_tokens", []) or []
         kwargs["additional_special_tokens"] += [
             code for code in FAIRSEQ_LANGUAGE_CODES if code not in kwargs["additional_special_tokens"]
         ]

src/transformers/models/mpnet/tokenization_mpnet.py

@@ -147,15 +147,15 @@ def __init__(
         strip_accents=None,
         **kwargs,
     ):
-        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
-        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
-        sep_token = AddedToken(sep_token, lstrip=False, rstrip=False) if isinstance(sep_token, str) else sep_token
-        cls_token = AddedToken(cls_token, lstrip=False, rstrip=False) if isinstance(cls_token, str) else cls_token
-        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
-        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        bos_token = AddedToken(bos_token, special=True) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, special=True) if isinstance(eos_token, str) else eos_token
+        sep_token = AddedToken(sep_token, special=True) if isinstance(sep_token, str) else sep_token
+        cls_token = AddedToken(cls_token, special=True) if isinstance(cls_token, str) else cls_token
+        unk_token = AddedToken(unk_token, special=True) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, special=True) if isinstance(pad_token, str) else pad_token
 
         # Mask token behave like a normal word, i.e. include the space before it
-        mask_token = AddedToken(mask_token, lstrip=True, rstrip=False) if isinstance(mask_token, str) else mask_token
+        mask_token = AddedToken(mask_token, lstrip=True, special=True) if isinstance(mask_token, str) else mask_token
 
         if not os.path.isfile(vocab_file):
             raise ValueError(
@@ -199,8 +199,9 @@ def vocab_size(self):
         return len(self.vocab)
 
     def get_vocab(self):
-        vocab = self.vocab.copy()
-        vocab.update(self.added_tokens_encoder)
+        # "<mask>" is part of the vocab, but was wrongfully added at a wrong index in the fast saved version
+        vocab = self.added_tokens_encoder.copy()
+        vocab.update(self.vocab)
         return vocab
 
     def _tokenize(self, text):