Add WhisperTokenizerFast (#21222)

* Add WhisperTokenizerFast

* Fixup

* Up

* Up

* Improve tests

* Update src/transformers/models/whisper/tokenization_whisper_fast.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Keep stride in whisper pipelien test

* Remove unknown token special case

* Reduce vocabulary size in tests

* Fix vocab size assertion

* Sync copied changes from WhisperTokenizer

* Skip pipeline tests

* Update assertion

* Remove Whisper tokenizer dependency on sentencepiece

* Format

---------

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

tests/models/whisper/test_modeling_tf_whisper.py

@@ -79,7 +79,7 @@ class TFWhisperModelTester:
         seq_length=60,
         is_training=True,
         use_labels=False,
-        vocab_size=99,
+        vocab_size=200,
         hidden_size=16,
         num_hidden_layers=2,
         num_attention_heads=4,

tests/models/whisper/test_modeling_whisper.py

@@ -96,7 +96,7 @@ class WhisperModelTester:
         seq_length=60,
         is_training=True,
         use_labels=False,
-        vocab_size=99,
+        vocab_size=200,
         hidden_size=16,
         num_hidden_layers=2,
         num_attention_heads=4,

tests/models/whisper/test_tokenization_whisper.py

@@ -14,7 +14,7 @@
 
 import unittest
 
-from transformers.models.whisper import WhisperTokenizer
+from transformers.models.whisper import WhisperTokenizer, WhisperTokenizerFast
 from transformers.testing_utils import slow
 
 from ...test_tokenization_common import TokenizerTesterMixin
@@ -31,7 +31,8 @@ NOTIMESTAMPS = 50363
 
 class WhisperTokenizerTest(TokenizerTesterMixin, unittest.TestCase):
     tokenizer_class = WhisperTokenizer
-    test_rust_tokenizer = False
+    rust_tokenizer_class = WhisperTokenizerFast
+    test_rust_tokenizer = True
     test_sentencepiece = False
     test_seq2seq = False
 
@@ -93,6 +94,17 @@ class WhisperTokenizerTest(TokenizerTesterMixin, unittest.TestCase):
     def test_tokenizer_slow_store_full_signature(self):
         pass
 
+    def test_tokenizer_fast_store_full_signature(self):
+        pass
+
+    def test_special_tokens_initialization(self):
+        # Whisper relies on specific additional special tokens, so we skip this
+        # general test. In particular, this test loads fast tokenizer from slow
+        # tokenizer, and the conversion uses prefix_tokens, where we reference
+        # additional special tokens by specific indices, hence overriding the
+        # list with less tokens leads to out of index error
+        pass
+
     @slow
     def test_tokenizer_integration(self):
         # fmt: off