Fix bloom add prefix space (#25652)

* properly support Sequence of pretokenizers

* actual fix

* make sure the fix works. Tests are not working for sure!

* hacky way

* add TODO

* update

* add a todo

* nits

* rename test

* nits

* rename test

src/transformers/models/bloom/tokenization_bloom_fast.py

@@ -135,7 +135,7 @@ class BloomTokenizerFast(PreTrainedTokenizerFast):
 
         if add_prefix_space:
             pre_tok_state = pre_tok_state.replace(b'"add_prefix_space":false', b'"add_prefix_space": true')
-            decoder_state = pre_tok_state.replace(b'"add_prefix_space":false', b'"add_prefix_space": true')
+            decoder_state = decoder_state.replace(b'"add_prefix_space":false', b'"add_prefix_space": true')
         self.backend_tokenizer.pre_tokenizer = pickle.loads(pre_tok_state)
         self.backend_tokenizer.decoder = pickle.loads(decoder_state)
 

tests/models/bloom/test_tokenization_bloom.py

@@ -133,3 +133,10 @@ class BloomTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
         # maximum sequence length of the positoonal embeddings.
         self.assertGreaterEqual(len(self.tokenizer_class.pretrained_vocab_files_map), 1)
         self.assertGreaterEqual(len(list(self.tokenizer_class.pretrained_vocab_files_map.values())[0]), 1)
+
+    def test_add_prefix_space_fast(self):
+        tokenizer_w_prefix = self.get_rust_tokenizer(add_prefix_space=True)
+        tokenizer_wo_prefix = self.get_rust_tokenizer(add_prefix_space=False)
+        tokens_w_prefix = tokenizer_w_prefix.tokenize("Hey")
+        tokens_wo_prefix = tokenizer_wo_prefix.tokenize("Hey")
+        self.assertNotEqual(tokens_w_prefix, tokens_wo_prefix)