Add support for FastSentencepieceTokenizer.

The implementation of the tokenizer is here: https://www.tensorflow.org/text/api_docs/python/text/FastSentencepieceTokenizer

This change should enable small student models, which use T5 as the backbone, to use the fast tokenizer, which for long sequences can slash several miliseconds of inference time on accelerators, since the tokenization happens on Device and not on Host.

PiperOrigin-RevId: 589615741

seqio/test_utils.py

@@ -1265,12 +1265,14 @@ def sentencepiece_vocab(
         sentencepiece_model_pb2.NormalizerSpec
     ] = None,
     reverse_extra_ids: bool = True,
+    use_fast_tokenizer: bool = False,
 ):
   return vocabularies.SentencePieceVocabulary(
       os.path.join(TEST_DATA_DIR, "sentencepiece", "sentencepiece.model"),
       extra_ids=extra_ids,
       normalizer_spec_overrides=normalizer_spec_overrides,
       reverse_extra_ids=reverse_extra_ids,
+      use_fast_tokenizer=use_fast_tokenizer,
   )
 
 

seqio/vocabularies.py

@@ -285,6 +285,9 @@ def __init__(
           sentencepiece_model_pb2.NormalizerSpec
       ] = None,
       reverse_extra_ids: bool = True,
+      # TODO(vladdoru): Flip this to True by default after we confirm there
+      # is no delta in the behavior of the 2 implementations.
+      use_fast_tokenizer: bool = False,
   ):
     """Create a SentencePieceVocabulary.
 
@@ -300,11 +303,14 @@ def __init__(
       reverse_extra_ids: if True, extra_ids are numbered in descending order, so
         the first extra_id has the highest number. This is done for
         compatibility with span_corruption mask generation in T5.
+      use_fast_tokenizer: use the tf_text fastsentencepiecetokenizer 
+        implementation which runs much faster.
     """
     self._sentencepiece_model_file = sentencepiece_model_file
     self._normalizer_spec_overrides = normalizer_spec_overrides
     self._reverse_extra_ids = reverse_extra_ids
     self._model: Optional[SentencePieceVocabulary._ModelContext] = None
+    self._use_fast_tokenizer = use_fast_tokenizer
 
     super().__init__(extra_ids=extra_ids)
 
@@ -436,6 +442,8 @@ def tokenizer(self) -> sentencepiece_processor.SentencePieceProcessor:
   @property
   def tf_tokenizer(self):
     """Instantiate and return a TF tokenizer."""
+    if self._use_fast_tokenizer:
+      return tf_text.FastSentencepieceTokenizer(model=self.sp_model)
     return tf_text.SentencepieceTokenizer(model=self.sp_model)
 
   @property

seqio/vocabularies_test.py

@@ -298,6 +298,20 @@ def test_extra_ids(self):
         test_tokens, tuple(vocab.encode_tf(test_string).numpy())
     )
 
+  def test_fast_tokenizer(self):
+    vocab = test_utils.sentencepiece_vocab(
+        extra_ids=10, use_fast_tokenizer=True)
+    self.assertEqual(36, vocab.vocab_size)
+    self.assertEqual("v", vocab.decode([25]))
+    test_string = "<extra_id_0> <extra_id_1> v <extra_id_9>"
+    test_tokens = (35, 34, 3, 25, 26)
+    self.assertEqual(test_string, vocab.decode(test_tokens))
+    self.assertEqual(test_string, _decode_tf(vocab, test_tokens))
+    self.assertSequenceEqual(test_tokens, vocab.encode(test_string))
+    self.assertSequenceEqual(
+        test_tokens, tuple(vocab.encode_tf(test_string).numpy())
+    )
+
   def test_force_repeated_whitespace_preservation(self):
     test_string = "a a  a   a"  # string with repeated whitespaces