Fix Qwen2Tokenizer (#29929)

qwen2: fixed tokens starting with # in slow tokenizer; add tests

Co-authored-by: jklj077 <17811943+jklj077@users.noreply.github.com>

src/transformers/models/qwen2/tokenization_qwen2.py

@@ -177,9 +177,9 @@ def __init__(
         self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
         bpe_merges = []
         with open(merges_file, encoding="utf-8") as merges_handle:
-            for line in merges_handle:
+            for i, line in enumerate(merges_handle):
                 line = line.strip()
-                if not line or line.startswith("#"):
+                if (i == 0 and line.startswith("#version:")) or not line:
                     continue
                 bpe_merges.append(tuple(line.split()))
         self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))

tests/models/qwen2/test_tokenization_qwen2.py

@@ -59,6 +59,8 @@ def setUp(self):
                 ";}",
                 ";}\u010a",
                 "\u00cf\u0135",
+                "\u0120#",
+                "##",
             ]
         )
 
@@ -75,6 +77,8 @@ def setUp(self):
             "; }",
             ";} \u010a",
             "\u00cf \u0135",
+            "\u0120 #",
+            "# #",
         ]
 
         self.special_tokens_map = {"eos_token": "<|endoftext|>"}
@@ -129,7 +133,7 @@ def test_python_full_tokenizer(self):
         self.assertListEqual(tokens, bpe_tokens)
 
         input_tokens = tokens
-        input_bpe_tokens = [75, 78, 86, 260, 259, 260, 220, 77, 68, 86, 260, 220, 15, 16, 15, 266, 268, 267]
+        input_bpe_tokens = [75, 78, 86, 260, 259, 260, 220, 77, 68, 86, 260, 220, 15, 16, 15, 266, 270, 267]
         self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
 
     @unittest.skip("We disable the test of pretokenization as it is not reversible.")
@@ -139,6 +143,11 @@ def test_pretokenized_inputs(self):
         # the results, by nature, should be different.
         pass
 
+    @unittest.skip("We disable the test of clean up tokenization spaces as it is not applicable.")
+    def test_clean_up_tokenization_spaces(self):
+        # it only tests bert-base-uncased and clean_up_tokenization_spaces is not applicable to this tokenizer
+        pass
+
     def test_nfc_normalization(self):
         # per https://unicode.org/faq/normalization.html, there are three characters whose normalization forms
         # under NFC, NFD, NFKC, and NFKD are all different
@@ -158,6 +167,16 @@ def test_nfc_normalization(self):
             tokenizer_output_string = tokenizer.backend_tokenizer.normalizer.normalize_str(input_string)
             self.assertEqual(tokenizer_output_string, output_string)
 
+    def test_slow_tokenizer_token_with_number_sign(self):
+        if not self.test_slow_tokenizer:
+            return
+
+        sequence = " ###"
+        token_ids = [268, 269]
+
+        tokenizer = self.get_tokenizer()
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokenizer.tokenize(sequence)), token_ids)
+
     def test_slow_tokenizer_decode_spaces_between_special_tokens_default(self):
         # Qwen2Tokenizer changes the default `spaces_between_special_tokens` in `decode` to False
         if not self.test_slow_tokenizer:
@@ -166,7 +185,7 @@ def test_slow_tokenizer_decode_spaces_between_special_tokens_default(self):
         # tokenizer has a special token: `"<|endfotext|>"` as eos, but it is not `legacy_added_tokens`
         # special tokens in `spaces_between_special_tokens` means spaces between `legacy_added_tokens`
         # that would be `"<|im_start|>"` and `"<|im_end|>"` in Qwen/Qwen2 Models
-        token_ids = [259, 260, 268, 269, 26]
+        token_ids = [259, 260, 270, 271, 26]
         sequence = " lower<|endoftext|><|im_start|>;"
         sequence_with_space = " lower<|endoftext|> <|im_start|> ;"