Fixing a pathological case for slow tokenizers (#14981)

* Fixing a pathological case for slow tokenizers

* Update src/transformers/tokenization_utils.py

src/transformers/tokenization_utils.py

@@ -131,7 +131,7 @@ def split(self, text: str) -> List[str]:
         # This is used by the lookahead which needs to skip over
         # some text where the full match exceeded the place in the initial
         # for loop
-        skip = None
+        skip = 0
         # Main loop, Giving this algorithm O(n) complexity
         for current, current_char in enumerate(text):
             if skip and current < skip:
@@ -175,6 +175,11 @@ def split(self, text: str) -> List[str]:
                             lookahead_index = current
                             end = current
                         next_char = text[lookahead_index] if lookahead_index < len(text) else None
+                        if "" in looktrie_pointer:
+                            start = lookstart
+                            end = lookahead_index
+                            skip = lookahead_index
+
                         while next_char in looktrie_pointer:
                             looktrie_pointer = looktrie_pointer[next_char]
                             lookahead_index += 1
@@ -219,7 +224,7 @@ def split(self, text: str) -> List[str]:
 
             # If this character is a starting character within the trie
             # start keeping track of this partial match.
-            if current_char in self.data:
+            if current >= skip and current_char in self.data:
                 states[current] = self.data[current_char]
 
         # We have a cut at the end with states.

tests/test_tokenization_common.py

@@ -3687,6 +3687,13 @@ def test_trie_suffix_tokens(self):
         trie.add("C")
         self.assertEqual(trie.split("ABC"), ["AB", "C"])
 
+    def test_trie_skip(self):
+        trie = Trie()
+        trie.add("ABC")
+        trie.add("B")
+        trie.add("CD")
+        self.assertEqual(trie.split("ABCD"), ["ABC", "D"])
+
     def test_cut_text_hardening(self):
         # Even if the offsets are wrong, we necessarily output correct string
         # parts.