fix train_new_from_iterator in the case of byte-level tokenizers

src/transformers/tokenization_utils_fast.py

@@ -21,6 +21,7 @@
 from collections import defaultdict
 from typing import Any, Dict, List, Optional, Tuple, Union
 
+import tokenizers.pre_tokenizers as pre_tokenizers_fast
 from tokenizers import Encoding as EncodingFast
 from tokenizers import Tokenizer as TokenizerFast
 from tokenizers.decoders import Decoder as DecoderFast
@@ -699,6 +700,8 @@ def train_new_from_iterator(
             kwargs["end_of_word_suffix"] = tokenizer_json["model"]["end_of_word_suffix"]
         if tokenizer_json["model"]["type"] == "Unigram" and unk_token is not None:
             kwargs["unk_token"] = unk_token
+        if tokenizer_json["pre_tokenizer"]["type"] == "ByteLevel":
+            kwargs["initial_alphabet"] = pre_tokenizers_fast.ByteLevel.alphabet()
 
         trainer_class = MODEL_TO_TRAINER_MAPPING[tokenizer_json["model"]["type"]]
         trainer = trainer_class(vocab_size=vocab_size, special_tokens=special_tokens, **kwargs)

tests/models/bart/test_modeling_bart.py

@@ -150,6 +150,7 @@ def get_config(self):
     def get_pipeline_config(self):
         config = self.get_config()
         config.max_position_embeddings = 100
+        config.vocab_size = 300
         return config
 
     def prepare_config_and_inputs_for_common(self):

tests/models/blenderbot/test_modeling_blenderbot.py

@@ -140,6 +140,7 @@ def get_config(self):
     def get_pipeline_config(self):
         config = self.get_config()
         config.max_position_embeddings = 100
+        config.vocab_size = 300
         return config
 
     def prepare_config_and_inputs_for_common(self):

tests/models/deberta/test_modeling_deberta.py

@@ -130,6 +130,11 @@ def get_config(self):
             pos_att_type=self.pos_att_type,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
     def check_loss_output(self, result):
         self.parent.assertListEqual(list(result.loss.size()), [])
 

tests/models/gpt2/test_modeling_gpt2.py

@@ -166,6 +166,11 @@ def get_config(
             reorder_and_upcast_attn=reorder_and_upcast_attn,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
     def prepare_config_and_inputs_for_decoder(self):
         (
             config,

tests/models/gpt_neo/test_modeling_gpt_neo.py

@@ -151,6 +151,11 @@ def get_config(self):
             attention_types=self.attention_types,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
     def prepare_config_and_inputs_for_decoder(self):
         (
             config,

tests/models/gptj/test_modeling_gptj.py

@@ -155,6 +155,11 @@ def get_config(self):
             rotary_dim=self.rotary_dim,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
     def prepare_config_and_inputs_for_decoder(self):
         (
             config,

tests/models/ibert/test_modeling_ibert.py

@@ -116,6 +116,11 @@ def get_config(self):
             quant_mode=True,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
     def create_and_check_model(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):

tests/models/led/test_modeling_led.py

@@ -163,6 +163,7 @@ def get_config(self):
     def get_pipeline_config(self):
         config = self.get_config()
         config.max_position_embeddings = 100
+        config.vocab_size = 300
         return config
 
     def prepare_config_and_inputs_for_common(self):

tests/models/longformer/test_modeling_longformer.py

@@ -113,6 +113,11 @@ def get_config(self):
             attention_window=self.attention_window,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
     def create_and_check_attention_mask_determinism(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):

tests/models/roberta/test_modeling_roberta.py

@@ -112,6 +112,11 @@ def get_config(self):
             initializer_range=self.initializer_range,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
     def prepare_config_and_inputs_for_decoder(self):
         (
             config,

tests/models/yoso/test_modeling_yoso.py

@@ -126,6 +126,11 @@ def get_config(self):
             initializer_range=self.initializer_range,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.vocab_size = 300
+        return config
+
     def prepare_config_and_inputs_for_decoder(self):
         (
             config,

tests/tokenization/test_tokenization_fast.py

@@ -39,6 +39,7 @@ def setUp(self):
         self.test_rust_tokenizer = True
 
         model_paths = ["robot-test/dummy-tokenizer-fast", "robot-test/dummy-tokenizer-wordlevel"]
+        self.bytelevel_bpe_model_name = "SaulLu/dummy-tokenizer-bytelevel-bpe"
 
         # Inclusion of 2 tokenizers to test different types of models (Unigram and WordLevel for the moment)
         self.tokenizers_list = [(PreTrainedTokenizerFast, model_path, {}) for model_path in model_paths]
@@ -99,6 +100,15 @@ def test_training_new_tokenizer_with_special_tokens_change(self):
                     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.assertEqual(encoding_ids, [64, 172, 253, 97, 245])
+
 
 @require_tokenizers
 class TokenizerVersioningTest(unittest.TestCase):