Add `from_slow` in fast tokenizers build and fixes some bugs (#9987)

src/transformers/models/albert/tokenization_albert.py

@@ -103,8 +103,9 @@ class AlbertTokenizer(PreTrainedTokenizer):
             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.
 
-    Attributes: sp_model (:obj:`SentencePieceProcessor`): The `SentencePiece` processor that is used for every
-    conversion (string, tokens and IDs).
+    Attributes:
+        sp_model (:obj:`SentencePieceProcessor`):
+            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/albert/tokenization_albert_fast.py

@@ -88,8 +88,11 @@ class AlbertTokenizerFast(PreTrainedTokenizerFast):
             Whether or not to keep accents when tokenizing.
         bos_token (:obj:`str`, `optional`, defaults to :obj:`"[CLS]"`):
             The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
-            .. note:: When building a sequence using special tokens, this is not the token that is used for the
-            beginning of sequence. The token used is the :obj:`cls_token`.
+
+            .. note::
+
+               When building a sequence using special tokens, this is not the token that is used for the beginning of
+               sequence. The token used is the :obj:`cls_token`.
         eos_token (:obj:`str`, `optional`, defaults to :obj:`"[SEP]"`):
             The end of sequence token. .. note:: When building a sequence using special tokens, this is not the token
             that is used for the end of sequence. The token used is the :obj:`sep_token`.
@@ -107,9 +110,7 @@ class AlbertTokenizerFast(PreTrainedTokenizerFast):
             instead of per-token classification). It is the first token of the sequence when built with special tokens.
         mask_token (:obj:`str`, `optional`, defaults to :obj:`"[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. Attributes:
-        sp_model (:obj:`SentencePieceProcessor`):
-            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
+            modeling. This is the token which the model will try to predict.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/auto/tokenization_auto.py

@@ -252,6 +252,7 @@ NO_CONFIG_TOKENIZER = [
     HerbertTokenizerFast,
     PhobertTokenizer,
     BarthezTokenizer,
+    BarthezTokenizerFast,
 ]
 
 

src/transformers/models/bart/tokenization_bart.py

@@ -38,11 +38,9 @@ class BartTokenizer(RobertaTokenizer):
     r"""
     Construct a BART tokenizer.
 
-    :class:`~transformers.BartTokenizer` is identical to :class:`~transformers.RobertaTokenizer` and adds a new
-    :meth:`~transformers.BartTokenizer.prepare_seq2seq_batch`
-
-    Refer to superclass :class:`~transformers.RobertaTokenizer` for usage examples and documentation concerning the
-    initialization parameters and other methods.
+    :class:`~transformers.BartTokenizer` is identical to :class:`~transformers.RobertaTokenizer`. Refer to superclass
+    :class:`~transformers.RobertaTokenizer` for usage examples and documentation concerning the initialization
+    parameters and other methods.
     """
     # merges and vocab same as Roberta
     max_model_input_sizes = {m: 1024 for m in _all_bart_models}

src/transformers/models/bart/tokenization_bart_fast.py

@@ -37,6 +37,13 @@ _all_bart_models = [
 
 
 class BartTokenizerFast(RobertaTokenizerFast):
+    r"""
+    Construct a "fast" BART tokenizer (backed by HuggingFace's `tokenizers` library).
+
+    :class:`~transformers.BartTokenizerFast` is identical to :class:`~transformers.RobertaTokenizerFast`. Refer to
+    superclass :class:`~transformers.RobertaTokenizerFast` for usage examples and documentation concerning the
+    initialization parameters and other methods.
+    """
     # merges and vocab same as Roberta
     max_model_input_sizes = {m: 1024 for m in _all_bart_models}
     pretrained_vocab_files_map = {

src/transformers/models/barthez/tokenization_barthez.py

@@ -90,8 +90,9 @@ class BarthezTokenizer(PreTrainedTokenizer):
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
 
-    Attributes: sp_model (:obj:`SentencePieceProcessor`): The `SentencePiece` processor that is used for every
-    conversion (string, tokens and IDs).
+    Attributes:
+        sp_model (:obj:`SentencePieceProcessor`):
+            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/barthez/tokenization_barthez_fast.py

@@ -98,9 +98,6 @@ class BarthezTokenizerFast(PreTrainedTokenizerFast):
             modeling. This is the token which the model will try to predict.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
-
-    Attributes: sp_model (:obj:`SentencePieceProcessor`): The `SentencePiece` processor that is used for every
-    conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/blenderbot/tokenization_blenderbot.py

@@ -36,7 +36,7 @@ class BlenderbotTokenizer(RobertaTokenizer):
     Construct a Blenderbot tokenizer.
 
     :class:`~transformers.Blenderbot` is nearly identical to :class:`~transformers.RobertaTokenizer` and runs
-    end-to-end tokenization: punctuation splitting and wordpiece. The only difference is that it doesnt add BOS token
+    end-to-end tokenization: punctuation splitting and wordpiece. The only difference is that it doesn't add BOS token
     to the beginning of sequences.
 
     Refer to superclass :class:`~transformers.RobertaTokenizer` for usage examples and documentation concerning

src/transformers/models/camembert/tokenization_camembert.py

@@ -93,8 +93,9 @@ class CamembertTokenizer(PreTrainedTokenizer):
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
 
-    Attributes: sp_model (:obj:`SentencePieceProcessor`): The `SentencePiece` processor that is used for every
-    conversion (string, tokens and IDs).
+    Attributes:
+        sp_model (:obj:`SentencePieceProcessor`):
+            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/camembert/tokenization_camembert_fast.py

@@ -101,10 +101,6 @@ class CamembertTokenizerFast(PreTrainedTokenizerFast):
             modeling. This is the token which the model will try to predict.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
-
-    Attributes:
-        sp_model (:obj:`SentencePieceProcessor`):
-            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/t5/tokenization_t5.py

@@ -92,6 +92,10 @@ class T5Tokenizer(PreTrainedTokenizer):
             <https://github.com/google-research/text-to-text-transfer-transformer/blob/9fd7b14a769417be33bc6c850f9598764913c833/t5/data/preprocessors.py#L2117>`__).
         additional_special_tokens (:obj:`List[str]`, `optional`):
             Additional special tokens used by the tokenizer.
+
+    Attributes:
+        sp_model (:obj:`SentencePieceProcessor`):
+            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/xlm_prophetnet/tokenization_xlm_prophetnet.py

@@ -97,8 +97,9 @@ class XLMProphetNetTokenizer(PreTrainedTokenizer):
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
 
-    Attributes: sp_model (:obj:`SentencePieceProcessor`): The `SentencePiece` processor that is used for every
-    conversion (string, tokens and IDs).
+    Attributes:
+        sp_model (:obj:`SentencePieceProcessor`):
+            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/xlm_roberta/tokenization_xlm_roberta.py

@@ -95,8 +95,9 @@ class XLMRobertaTokenizer(PreTrainedTokenizer):
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
 
-    Attributes: sp_model (:obj:`SentencePieceProcessor`): The `SentencePiece` processor that is used for every
-    conversion (string, tokens and IDs).
+    Attributes:
+        sp_model (:obj:`SentencePieceProcessor`):
+            The `SentencePiece` processor that is used for every conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/models/xlm_roberta/tokenization_xlm_roberta_fast.py

@@ -106,9 +106,6 @@ class XLMRobertaTokenizerFast(PreTrainedTokenizerFast):
             modeling. This is the token which the model will try to predict.
         additional_special_tokens (:obj:`List[str]`, `optional`, defaults to :obj:`["<s>NOTUSED", "</s>NOTUSED"]`):
             Additional special tokens used by the tokenizer.
-
-    Attributes: sp_model (:obj:`SentencePieceProcessor`): The `SentencePiece` processor that is used for every
-    conversion (string, tokens and IDs).
     """
 
     vocab_files_names = VOCAB_FILES_NAMES

src/transformers/tokenization_utils_base.py

@@ -1793,12 +1793,11 @@ class PreTrainedTokenizerBase(SpecialTokensMixin):
     def _from_pretrained(
         cls, resolved_vocab_files, pretrained_model_name_or_path, init_configuration, *init_inputs, **kwargs
     ):
-        # We instantiate fast tokenizers based on a slow tokenizer for now
-        # In the future we can also use a direct way based on saving/instantiating
-        # tokenizer's Tokenizer directly from it's serialization JSON
-        if (
-            "tokenizer_file" not in resolved_vocab_files or resolved_vocab_files["tokenizer_file"] is None
-        ) and cls.slow_tokenizer_class is not None:
+        # We instantiate fast tokenizers based on a slow tokenizer if we don't have access to the tokenizer.json
+        # file or if `from_slow` is set to True.
+        from_slow = kwargs.get("from_slow", False)
+        has_tokenizer_file = resolved_vocab_files.get("tokenizer_file", None) is not None
+        if (from_slow or not has_tokenizer_file) and cls.slow_tokenizer_class is not None:
             slow_tokenizer = (cls.slow_tokenizer_class)._from_pretrained(
                 copy.deepcopy(resolved_vocab_files),
                 pretrained_model_name_or_path,

src/transformers/tokenization_utils_fast.py

@@ -80,8 +80,15 @@ class PreTrainedTokenizerFast(PreTrainedTokenizerBase):
     def __init__(self, *args, **kwargs):
         slow_tokenizer = kwargs.pop("__slow_tokenizer", None)
         fast_tokenizer_file = kwargs.pop("tokenizer_file", None)
+        from_slow = kwargs.pop("from_slow", False)
 
-        if fast_tokenizer_file is not None:
+        if from_slow and slow_tokenizer is None and self.slow_tokenizer_class is None:
+            raise ValueError(
+                "Cannot instantiate this tokenizer from a slow version. If it's based on sentencepiece, make sure you "
+                "have sentencepiece installed."
+            )
+
+        if fast_tokenizer_file is not None and not from_slow:
             # We have a serialization from tokenizers which let us directly build the backend
             fast_tokenizer = TokenizerFast.from_file(fast_tokenizer_file)
         elif slow_tokenizer is not None: