Fix slow GemmaTokenizer and improve SPM slow -> fast conversion process (#32191)

* Remove user-defined tokens which can be obtained through merges

* Remove debug line

* formatting

* Refactor spm slow -> fast converter

* revert unnecessary refactor

* set comprehension

* remove test files

* Use `vocab_scores`

* Always replace spiece underline with space in decode

* we no longer need token filtering

* Add save fast load slow unit test

* Remove tokenizers version check

* Remove duplicate code

* Make `<start_of_turn>` and `<end_of_turn>` special tokens

* Bias merge priority with length if score is the same

* Add unit test for merge priority

* CI

src/transformers/convert_slow_tokenizer.py

@@ -53,6 +53,25 @@ def _get_prepend_scheme(add_prefix_space: bool, original_tokenizer) -> str:
     return prepend_scheme
 
 
+def generate_merges(vocab, vocab_scores):
+    reverse = vocab_scores is not None
+    vocab_scores = dict(vocab_scores) if reverse else vocab
+
+    merges = []
+    for merge, piece_score in vocab_scores.items():
+        local = []
+        for index in range(1, len(merge)):
+            piece_l, piece_r = merge[:index], merge[index:]
+            if piece_l in vocab and piece_r in vocab:
+                local.append((piece_l, piece_r, piece_score))
+        local = sorted(local, key=lambda x: (vocab[x[0]], vocab[x[1]]))
+        merges.extend(local)
+
+    merges = sorted(merges, key=lambda val: (val[2], len(val[0]), len(val[1])), reverse=reverse)
+    merges = [(val[0], val[1]) for val in merges]
+    return merges
+
+
 class SentencePieceExtractor:
     """
     Extractor implementation for SentencePiece trained models. https://github.com/google/sentencepiece
@@ -73,24 +92,8 @@ class SentencePieceExtractor:
         sp = self.sp
         vocab = {sp.id_to_piece(index): index for index in range(sp.GetPieceSize())}
 
-        if vocab_scores is not None:
-            vocab_scores, reverse = dict(vocab_scores), True
-        else:
-            vocab_scores, reverse = vocab, False
-
-        # Merges
-        merges = []
-        for merge, piece_score in vocab_scores.items():
-            local = []
-            for index in range(1, len(merge)):
-                piece_l, piece_r = merge[:index], merge[index:]
-                if piece_l in vocab and piece_r in vocab:
-                    local.append((piece_l, piece_r, piece_score))
-            local = sorted(local, key=lambda x: (vocab[x[0]], vocab[x[1]]))
-            merges.extend(local)
+        merges = generate_merges(vocab, vocab_scores)
 
-        merges = sorted(merges, key=lambda val: val[2], reverse=reverse)
-        merges = [(val[0], val[1]) for val in merges]
         return vocab, merges
 
 
@@ -107,24 +110,7 @@ class GemmaSentencePieceExtractor(SentencePieceExtractor):
         # "<0x09>" is the bytefallback for `\t`
         vocab["\t"] = vocab.get("<0x09>")
 
-        if vocab_scores is not None:
-            vocab_scores, reverse = dict(vocab_scores), True
-        else:
-            vocab_scores, reverse = vocab, False
-
-        # Merges
-        merges = []
-        for merge, piece_score in vocab_scores.items():
-            local = []
-            for index in range(1, len(merge)):
-                piece_l, piece_r = merge[:index], merge[index:]
-                if piece_l in vocab and piece_r in vocab:
-                    local.append((piece_l, piece_r, piece_score))
-            local = sorted(local, key=lambda x: (vocab[x[0]], vocab[x[1]]))
-            merges.extend(local)
-
-        merges = sorted(merges, key=lambda val: val[2], reverse=reverse)
-        merges = [(val[0], val[1]) for val in merges]
+        merges = generate_merges(vocab, vocab_scores)
         return vocab, merges
 
 
@@ -544,6 +530,10 @@ class DebertaConverter(Converter):
 
 
 class SpmConverter(Converter):
+    handle_byte_fallback = False
+    SpmExtractor = SentencePieceExtractor
+    special_tokens = {}
+
     def __init__(self, *args):
         requires_backends(self, "protobuf")
 
@@ -557,8 +547,7 @@ class SpmConverter(Converter):
             m.ParseFromString(f.read())
         self.proto = m
 
-        if self.proto.trainer_spec.byte_fallback:
-            if not getattr(self, "handle_byte_fallback", None):
+        if self.proto.trainer_spec.byte_fallback and not self.handle_byte_fallback:
             warnings.warn(
                 "The sentencepiece tokenizer that you are converting to a fast tokenizer uses the byte fallback option"
                 " which is not implemented in the fast tokenizers. In practice this means that the fast version of the"
@@ -575,12 +564,18 @@ class SpmConverter(Converter):
     def tokenizer(self, proto):
         model_type = proto.trainer_spec.model_type
         vocab_scores = self.vocab(proto)
-        unk_id = self.unk_id(proto)
 
         if model_type == 1:
-            tokenizer = Tokenizer(Unigram(vocab_scores, unk_id))
+            tokenizer = Tokenizer(
+                Unigram(
+                    vocab_scores,
+                    unk_id=self.unk_id(proto),
+                    byte_fallback=self.handle_byte_fallback,
+                )
+            )
+
         elif model_type == 2:
-            _, merges = SentencePieceExtractor(self.original_tokenizer.vocab_file).extract()
+            _, merges = self.SpmExtractor(self.original_tokenizer.vocab_file).extract(vocab_scores)
             bpe_vocab = {word: i for i, (word, score) in enumerate(vocab_scores)}
             tokenizer = Tokenizer(
                 BPE(
@@ -588,13 +583,53 @@ class SpmConverter(Converter):
                     merges,
                     unk_token=proto.trainer_spec.unk_piece,
                     fuse_unk=True,
+                    byte_fallback=self.handle_byte_fallback,
+                    dropout=None,
                 )
             )
+
         else:
             raise Exception(
                 "You're trying to run a `Unigram` model but you're file was trained with a different algorithm"
             )
 
+        # control tokens are special
+        # user defined symbols are not
+        # both user and control tokens are AddedTokens
+        # Add user defined symbols (type == 4) from sentencepiece (https://github.com/google/sentencepiece/blob/6225e08edb2577757163b3f5dbba4c0b670ef445/src/sentencepiece_model.proto#L299C29-L299C33)
+        spm_added_tokens = [
+            (id, p.piece, p.type == 3 or p.piece in self.special_tokens)
+            for id, p in enumerate(proto.pieces)
+            if p.type in [3, 4]
+        ]
+        tokens_to_add = [
+            AddedToken(token, normalized=False, special=special)
+            for id, token, special in sorted(spm_added_tokens, key=lambda x: x[0])
+        ]
+
+        if len(tokens_to_add) > 0:
+            # super hack: if a token.special is set, tokenizer ignores it for now so FIXME @ArthurZ
+            # Accumulate added tokens into batches of special/non-special tokens, because calling add_tokens() for
+            # individual tokens would repeatedly rebuild a trie, which can be slow.
+            is_last_special = None
+            tokens = []
+            for token in tokens_to_add:
+                is_special = token.special
+                if is_last_special is None or is_last_special == is_special:
+                    tokens.append(token)
+                else:
+                    if is_last_special:
+                        tokenizer.add_special_tokens(tokens)
+                    else:
+                        tokenizer.add_tokens(tokens)
+                    tokens = [token]
+                is_last_special = is_special
+            if tokens:
+                if is_last_special:
+                    tokenizer.add_special_tokens(tokens)
+                else:
+                    tokenizer.add_tokens(tokens)
+
         return tokenizer
 
     def normalizer(self, proto):
@@ -622,40 +657,6 @@ class SpmConverter(Converter):
     def converted(self) -> Tokenizer:
         tokenizer = self.tokenizer(self.proto)
 
-        # control tokens are special
-        # user defined symbols are not
-        # both user and control tokens are AddedTokens
-        # Add user defined symbols (type == 4) from sentnecepiece (https://github.com/google/sentencepiece/blob/6225e08edb2577757163b3f5dbba4c0b670ef445/src/sentencepiece_model.proto#L299C29-L299C33)
-
-        tokens_to_add = {
-            id: AddedToken(token, normalized=False, special=special)
-            for id, token, special in [
-                (id, p.piece, p.type == 3) for id, p in enumerate(self.proto.pieces) if p.type in [3, 4]
-            ]
-        }
-        tokens_to_add = [k for _, k in sorted(tokens_to_add.items(), key=lambda x: x[0])]
-        if len(tokens_to_add) > 0:
-            # super hack: if a token.special is set, tokenizer ignores it for now so FIXME @ArthurZ
-            # Accumulate added tokens into batches of special/non-special tokens, because calling add_tokens() for
-            # individual tokens would repeatedly rebuild a trie, which can be slow.
-            is_last_special = None
-            tokens = []
-            for token in tokens_to_add:
-                is_special = token.special
-                if is_last_special is None or is_last_special == is_special:
-                    tokens.append(token)
-                else:
-                    if is_last_special:
-                        tokenizer.add_special_tokens(tokens)
-                    else:
-                        tokenizer.add_tokens(tokens)
-                    tokens = [token]
-                is_last_special = is_special
-            if tokens:
-                if is_last_special:
-                    tokenizer.add_special_tokens(tokens)
-                else:
-                    tokenizer.add_tokens(tokens)
         # Tokenizer assemble
         normalizer = self.normalizer(self.proto)
         if normalizer is not None:
@@ -1283,6 +1284,9 @@ class XGLMConverter(SpmConverter):
 
 class GemmaConvert(SpmConverter):
     handle_byte_fallback = True
+    SpmExtractor = GemmaSentencePieceExtractor
+    # start and end of turn tokens must be marked as special
+    special_tokens = {"<start_of_turn>", "<end_of_turn>"}
 
     """"
     split_by_unicode_script: true
@@ -1327,45 +1331,6 @@ class GemmaConvert(SpmConverter):
             ]
         )
 
-    def tokenizer(self, proto):
-        model_type = proto.trainer_spec.model_type
-        vocab_scores = self.vocab(proto)
-        if model_type == 1:
-            import tokenizers
-
-            if version.parse(tokenizers.__version__) < version.parse("0.14.0"):
-                tokenizer = Tokenizer(Unigram(vocab_scores, 0))
-            else:
-                tokenizer = Tokenizer(Unigram(vocab_scores, 0, byte_fallback=True))
-
-        elif model_type == 2:
-            _, merges = GemmaSentencePieceExtractor(self.original_tokenizer.vocab_file).extract(vocab_scores)
-            bpe_vocab = {word: i for i, (word, _score) in enumerate(vocab_scores)}
-
-            tokenizer = Tokenizer(
-                BPE(
-                    bpe_vocab,
-                    merges,
-                    unk_token=proto.trainer_spec.unk_piece,
-                    fuse_unk=True,
-                    byte_fallback=True,
-                    dropout=None,
-                )
-            )
-            tokenizer.add_special_tokens(
-                [
-                    AddedToken("<pad>", normalized=False, special=True),
-                    AddedToken("<eos>", normalized=False, special=True),
-                    AddedToken("<bos>", normalized=False, special=True),
-                    AddedToken("<unk>", normalized=False, special=True),
-                ]
-            )
-        else:
-            raise Exception(
-                "You're trying to run a `Unigram` model but you're file was trained with a different algorithm"
-            )
-        return tokenizer
-
 
 class LlamaConverter(SpmConverter):
     handle_byte_fallback = True
@@ -1393,37 +1358,6 @@ class LlamaConverter(SpmConverter):
             sequence += [decoders.Strip(content=" ", left=1)]
         return decoders.Sequence(sequence)
 
-    def tokenizer(self, proto):
-        model_type = proto.trainer_spec.model_type
-        vocab_scores = self.vocab(proto)
-        if model_type == 1:
-            import tokenizers
-
-            if version.parse(tokenizers.__version__) < version.parse("0.14.0"):
-                tokenizer = Tokenizer(Unigram(vocab_scores, 0))
-            else:
-                tokenizer = Tokenizer(Unigram(vocab_scores, 0, byte_fallback=True))
-
-        elif model_type == 2:
-            _, merges = SentencePieceExtractor(self.original_tokenizer.vocab_file).extract(vocab_scores)
-            bpe_vocab = {word: i for i, (word, _score) in enumerate(vocab_scores)}
-            tokenizer = Tokenizer(
-                BPE(bpe_vocab, merges, unk_token=proto.trainer_spec.unk_piece, fuse_unk=True, byte_fallback=True)
-            )
-            tokenizer.add_special_tokens(
-                [
-                    AddedToken(self.original_tokenizer.convert_ids_to_tokens(0), normalized=False, special=True),
-                    AddedToken(self.original_tokenizer.convert_ids_to_tokens(1), normalized=False, special=True),
-                    AddedToken(self.original_tokenizer.convert_ids_to_tokens(2), normalized=False, special=True),
-                ]
-            )
-        else:
-            raise Exception(
-                "You're trying to run a `Unigram` model but you're file was trained with a different algorithm"
-            )
-
-        return tokenizer
-
     def normalizer(self, proto):
         if getattr(self.original_tokenizer, "legacy", True):
             sequence = []

src/transformers/models/gemma/tokenization_gemma.py

@@ -198,7 +198,7 @@ class GemmaTokenizer(PreTrainedTokenizer):
         else:
             sub_texts = "".join(sub_texts)
 
-        return sub_texts
+        return sub_texts.replace(SPIECE_UNDERLINE, " ")
 
     def convert_tokens_to_string(self, tokens):
         """Converts a sequence of tokens (string) in a single string."""

tests/models/gemma/test_tokenization_gemma.py

@@ -222,6 +222,17 @@ class GemmaIntegrationTest(unittest.TestCase):
         self.tokenizer.add_eos_token = False
         self.rust_tokenizer.add_eos_token = False
 
+    def test_fast_merge_priority(self):
+        slow_tokenizer = self.tokenizer
+        fast_tokenizer = self.rust_tokenizer
+        text = "                                               "
+        target = [168, 153]
+        slow = slow_tokenizer.encode(text, add_special_tokens=False)
+        assert slow == target
+
+        fast = fast_tokenizer.encode(text, add_special_tokens=False)
+        assert fast == target
+
     @unittest.skip(reason="Not super important and always failing. Let's skip it")
     @slow
     def test_conversion(self):
@@ -442,6 +453,30 @@ class GemmaIntegrationTest(unittest.TestCase):
         for tokenized_chat, expected_tokens in zip(tokenized_chats, expected_tokens):
             self.assertListEqual(tokenized_chat, expected_tokens)
 
+    def test_save_fast_load_slow(self):
+        # Ensure that we can save a fast tokenizer and load it as a slow tokenizer
+        slow_tokenizer = self.tokenizer
+        text = "a  "
+        target_encoded = [2, 235250, 139]
+        slow = slow_tokenizer.encode(text, add_special_tokens=True)
+        assert slow == target_encoded
+
+        slow_decoded = slow_tokenizer.decode(slow, skip_special_tokens=True)
+        assert slow_decoded == text
+
+        with tempfile.TemporaryDirectory() as dirname:
+            # Save fast tokenizer
+            self.rust_tokenizer.save_pretrained(dirname)
+
+            # Load slow tokenizer with fast files present in the directory
+            slow_tokenizer_from_fast = GemmaTokenizer.from_pretrained(dirname)
+
+        slow_from_fast = slow_tokenizer_from_fast.encode(text, add_special_tokens=True)
+        assert slow_from_fast == target_encoded
+
+        slow_from_fast_decoded = slow_tokenizer_from_fast.decode(slow, skip_special_tokens=True)
+        assert slow_from_fast_decoded == text
+
 
 @require_sentencepiece
 @require_tokenizers