transformers

1
# coding=utf-8
2
# Copyright 2024 The HuggingFace Team. All rights reserved.
3
#
4
# Licensed under the Apache License, Version 2.0 (the "License");
5
# you may not use this file except in compliance with the License.
6
# You may obtain a copy of the License at
7
#
8
#     http://www.apache.org/licenses/LICENSE-2.0
9
#
10
# Unless required by applicable law or agreed to in writing, software
11
# distributed under the License is distributed on an "AS IS" BASIS,
12
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
# See the License for the specific language governing permissions and
14
# limitations under the License.
15

16
import os
17
import tempfile
18
import unittest
19

20
from datasets import load_dataset
21

22
from transformers import (
23
    AddedToken,
24
    GemmaTokenizer,
25
    GemmaTokenizerFast,
26
    is_torch_available,
27
)
28
from transformers.convert_slow_tokenizer import convert_slow_tokenizer
29
from transformers.testing_utils import (
30
    get_tests_dir,
31
    nested_simplify,
32
    require_jinja,
33
    require_sentencepiece,
34
    require_tokenizers,
35
    require_torch,
36
    slow,
37
)
38

39
from ...test_tokenization_common import TokenizerTesterMixin
40

41

42
SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
43

44

45
if is_torch_available():
46
    pass
47

48

49
@require_sentencepiece
50
@require_tokenizers
51
class GemmaTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
52
    tokenizer_class = GemmaTokenizer
53
    rust_tokenizer_class = GemmaTokenizerFast
54

55
    test_rust_tokenizer = False
56
    test_sentencepiece = True
57
    from_pretrained_kwargs = {}
58

59
    def setUp(self):
60
        super().setUp()
61
        # We have a SentencePiece fixture for testing
62
        tokenizer = GemmaTokenizer(SAMPLE_VOCAB, keep_accents=True)
63
        tokenizer.pad_token = tokenizer.eos_token
64
        tokenizer.save_pretrained(self.tmpdirname)
65

66
    @require_torch
67
    def test_batch_tokenization(self):
68
        if not self.test_seq2seq:
69
            return
70

71
        tokenizers = self.get_tokenizers()
72
        for tokenizer in tokenizers:
73
            with self.subTest(f"{tokenizer.__class__.__name__}"):
74
                # Longer text that will definitely require truncation.
75
                text = [
76
                    " UN Chief Says There Is No Military Solution in Syria",
77
                    " Secretary-General Ban Ki-moon says his response to Russia's stepped up military support for"
78
                    " Syria is that 'there is no military solution' to the nearly five-year conflict and more weapons"
79
                    " will only worsen the violence and misery for millions of people.",
80
                ]
81
                try:
82
                    batch = tokenizer(
83
                        text=text,
84
                        max_length=3,
85
                        max_target_length=10,
86
                        return_tensors="pt",
87
                    )
88
                except NotImplementedError:
89
                    return
90
                self.assertEqual(batch.input_ids.shape[1], 3)
91
                # max_target_length will default to max_length if not specified
92
                batch = tokenizer(text, max_length=3, return_tensors="pt")
93
                self.assertEqual(batch.input_ids.shape[1], 3)
94

95
                batch_encoder_only = tokenizer(text=text, max_length=3, max_target_length=10, return_tensors="pt")
96
                self.assertEqual(batch_encoder_only.input_ids.shape[1], 3)
97
                self.assertEqual(batch_encoder_only.attention_mask.shape[1], 3)
98
                self.assertNotIn("decoder_input_ids", batch_encoder_only)
99

100
    @unittest.skip("Unfortunately way too slow to build a BPE with SentencePiece.")
101
    def test_save_slow_from_fast_and_reload_fast(self):
102
        pass
103

104
    def test_special_tokens_initialization(self):
105
        for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
106
            with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
107
                added_tokens = [AddedToken("<special>", lstrip=True)]
108

109
                tokenizer_r = self.rust_tokenizer_class.from_pretrained(
110
                    pretrained_name, additional_special_tokens=added_tokens, **kwargs
111
                )
112
                r_output = tokenizer_r.encode("Hey this is a <special> token")
113

114
                special_token_id = tokenizer_r.encode("<special>", add_special_tokens=False)[0]
115

116
                self.assertTrue(special_token_id in r_output)
117

118
                if self.test_slow_tokenizer:
119
                    tokenizer_cr = self.rust_tokenizer_class.from_pretrained(
120
                        pretrained_name,
121
                        additional_special_tokens=added_tokens,
122
                        **kwargs,  # , from_slow=True <- unfortunately too slow to convert
123
                    )
124
                    tokenizer_p = self.tokenizer_class.from_pretrained(
125
                        pretrained_name, additional_special_tokens=added_tokens, **kwargs
126
                    )
127

128
                    p_output = tokenizer_p.encode("Hey this is a <special> token")
129

130
                    cr_output = tokenizer_cr.encode("Hey this is a <special> token")
131

132
                    self.assertEqual(p_output, r_output)
133
                    self.assertEqual(cr_output, r_output)
134
                    self.assertTrue(special_token_id in p_output)
135
                    self.assertTrue(special_token_id in cr_output)
136

137
    @slow
138
    def test_tokenizer_integration(self):
139
        expected_encoding =  {'input_ids': [[2, 158434, 591, 84193, 3836, 685, 6599, 31223, 235290, 140247, 578, 6599, 31223, 235290, 145139, 235290, 3491, 235275, 6572, 3311, 235290, 38197, 109959, 591, 25894, 235269, 162174, 235290, 235284, 235269, 1791, 6362, 12481, 235269, 1576, 18622, 235269, 2900, 1136, 86684, 235269, 29092, 4632, 16994, 604, 13146, 14944, 40371, 591, 19700, 235327, 235275, 578, 13146, 14944, 25511, 591, 235300, 12474, 235275, 675, 1163, 235248, 235304, 235284, 235340, 229903, 5377, 575, 235248, 235274, 235276, 235276, 235340, 17044, 578, 5271, 1061, 118345, 1865, 125247, 235269, 8745, 111226, 578, 176888, 235265], [2, 25894, 603, 6869, 577, 953, 235290, 8297, 5271, 209099, 41642, 774, 748, 78253, 2793, 731, 51506, 34346, 611, 2145, 2731, 578, 1833, 4807, 575, 832, 16630, 235265], [2, 651, 4320, 8426, 25341, 36271, 1163, 573, 27894, 5929, 235265]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]}  # fmt: skip
140
        self.tokenizer_integration_test_util(
141
            expected_encoding=expected_encoding,
142
            model_name="hf-internal-testing/dummy-gemma",
143
            revision="",
144
            padding=False,
145
        )
146

147
    @unittest.skip("worker 'gw4' crashed on CI, passing locally.")
148
    def test_pickle_subword_regularization_tokenizer(self):
149
        pass
150

151
    @unittest.skip("worker 'gw4' crashed on CI, passing locally.")
152
    def test_subword_regularization_tokenizer(self):
153
        pass
154

155
    @unittest.skip("This test will be removed from main @LysandreJik")
156
    def test_pretrained_model_lists(self):
157
        pass
158

159
    @unittest.skip("Skipping")
160
    def test_torch_encode_plus_sent_to_model(self):
161
        pass
162

163

164
@require_torch
165
@require_sentencepiece
166
@require_tokenizers
167
class GemmaIntegrationTest(unittest.TestCase):
168
    @classmethod
169
    def setUpClass(cls):
170
        checkpoint_name = "hf-internal-testing/dummy-gemma"
171
        cls.tokenizer: GemmaTokenizer = GemmaTokenizer.from_pretrained(
172
            checkpoint_name, eos_token="<s>"
173
        )  # add this token
174
        cls.rust_tokenizer = GemmaTokenizerFast.from_pretrained(
175
            checkpoint_name, eos_token="<s>", from_slow=True
176
        )  # add this token
177
        return cls
178

179
    @require_torch
180
    def integration_tests(self):
181
        inputs = self.tokenizer(
182
            ["The following string should be properly encoded: Hello.", "But ird and ปี   ird   ด"],
183
            return_tensors="pt",
184
        )
185

186
        self.assertEqual(
187
            nested_simplify(inputs),
188
            {
189
                "input_ids": [
190
                    [2, 450, 1494, 1347, 881, 367, 6284, 18511, 29901, 15043, 29889],
191
                    [2, 1205, 29871, 1823, 322, 29871, 31010, 30691, 1678, 1823, 1678, 30718],
192
                ],
193
                "attention_mask": [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]],
194
            },
195
        )
196

197
    def test_fast_special_tokens(self):
198
        slow_tokenizer = self.tokenizer
199
        fast_tokenizer = self.rust_tokenizer
200
        slow = slow_tokenizer.encode("A sample test", add_special_tokens=True)
201
        assert slow == [2, 235280, 6453, 2121]
202

203
        fast_tokenizer.add_eos_token = False
204
        fast = fast_tokenizer.encode("A sample test", add_special_tokens=True)
205
        assert fast == [2, 235280, 6453, 2121]
206

207
        fast_tokenizer.add_eos_token = True
208
        fast = fast_tokenizer.encode("A sample test", add_special_tokens=True)
209
        assert fast == [2, 235280, 6453, 2121, 204]
210

211
        slow_tokenizer.add_eos_token = True
212
        slow = slow_tokenizer.encode("A sample test", add_special_tokens=True)
213
        assert slow == [2, 235280, 6453, 2121, 204]
214

215
        self.tokenizer.add_eos_token = False
216
        self.rust_tokenizer.add_eos_token = False
217

218
    @unittest.skip("Not super important and always failing. Let's skip it")
219
    @slow
220
    def test_conversion(self):
221
        # This is excruciatingly slow since it has to recreate the entire merge
222
        # list from the original vocabulary in spm
223
        self.rust_tokenizer.save_pretrained("./out")
224
        with tempfile.TemporaryDirectory() as dirname:
225
            self.rust_tokenizer.save_pretrained(dirname)
226

227
            with open(os.path.join(dirname, "tokenizer.json"), "r") as f:
228
                old_serialized = f.read()
229

230
        new_tokenizer = convert_slow_tokenizer(self.tokenizer)
231
        with tempfile.NamedTemporaryFile() as f:
232
            new_tokenizer.save(f.name)
233
            # Re-opening since `f` is in bytes.
234
            new_serialized = open(f.name, "r").read()
235
            with open("out_tokenizer.json", "w") as g:
236
                g.write(new_serialized)
237

238
            self.assertEqual(old_serialized, new_serialized)
239

240
    def test_simple_encode_decode(self):
241
        pyth_tokenizer = self.tokenizer
242
        rust_tokenizer = self.rust_tokenizer
243

244
        self.tokenizer.add_eos_token = False
245
        self.rust_tokenizer.add_eos_token = False
246

247
        self.assertEqual(pyth_tokenizer.encode("This is a test"), [2, 1596, 603, 476, 2121])
248
        self.assertEqual(rust_tokenizer.encode("This is a test"), [2, 1596, 603, 476, 2121])
249
        self.assertEqual(pyth_tokenizer.decode([2, 1596, 603, 476, 2121], skip_special_tokens=True), "This is a test")
250
        self.assertEqual(rust_tokenizer.decode([2, 1596, 603, 476, 2121], skip_special_tokens=True), "This is a test")
251

252
        # bytefallback showcase
253
        self.assertEqual(pyth_tokenizer.encode("生活的真谛是"), [2, 122182, 235710, 245467, 235427] )  # fmt: skip
254
        self.assertEqual(rust_tokenizer.encode("生活的真谛是"), [2, 122182, 235710, 245467, 235427] )  # fmt: skip
255
        self.assertEqual(
256
            pyth_tokenizer.decode([2, 122182, 235710, 245467, 235427], skip_special_tokens=True),
257
            "生活的真谛是",
258
        )
259
        self.assertEqual(
260
            rust_tokenizer.decode([2, 122182, 235710, 245467, 235427], skip_special_tokens=True),
261
            "生活的真谛是",
262
        )
263

264
        # Inner spaces showcase
265
        self.assertEqual(pyth_tokenizer.encode("Hi  Hello"), [2, 2151, 139, 4521])
266
        self.assertEqual(rust_tokenizer.encode("Hi  Hello"), [2, 2151, 139, 4521])
267
        self.assertEqual(pyth_tokenizer.decode([2, 2151, 139, 4521], skip_special_tokens=True), "Hi  Hello")
268
        self.assertEqual(rust_tokenizer.decode([2, 2151, 139, 4521], skip_special_tokens=True), "Hi  Hello")
269

270
        self.assertEqual(pyth_tokenizer.encode("Hi   Hello"), [2, 2151, 140, 4521])
271
        self.assertEqual(rust_tokenizer.encode("Hi   Hello"), [2, 2151, 140, 4521])
272
        self.assertEqual(pyth_tokenizer.decode([2, 2151, 140, 4521], skip_special_tokens=True), "Hi   Hello")
273
        self.assertEqual(rust_tokenizer.decode([2, 2151, 140, 4521], skip_special_tokens=True), "Hi   Hello")
274

275
        self.assertEqual(pyth_tokenizer.encode(""), [2])
276
        self.assertEqual(rust_tokenizer.encode(""), [2])
277

278
        self.assertEqual(pyth_tokenizer.encode(" "), [2, 235248])
279
        self.assertEqual(rust_tokenizer.encode(" "), [2, 235248])
280

281
        self.assertEqual(pyth_tokenizer.encode("  "), [2, 139])
282
        self.assertEqual(rust_tokenizer.encode("  "), [2, 139])
283

284
        self.assertEqual(pyth_tokenizer.encode(" Hello"), [2, 25957])
285
        self.assertEqual(rust_tokenizer.encode(" Hello"), [2, 25957])
286

287
    def test_no_differences_decode(self):
288
        self.tokenizer.add_eos_token = False
289
        self.rust_tokenizer.add_eos_token = False
290
        pyth_tokenizer = self.tokenizer
291
        rust_tokenizer = self.rust_tokenizer
292

293
        self.assertEqual(pyth_tokenizer.decode([869]), "og")
294
        self.assertEqual(rust_tokenizer.decode([869]), "og")
295

296
        self.assertEqual(pyth_tokenizer.decode([30112, 869]), " expenditureog")
297
        self.assertEqual(rust_tokenizer.decode([30112, 869]), " expenditureog")
298

299
    def test_no_differences_special_tokens(self):
300
        pyth_tokenizer = self.tokenizer
301
        rust_tokenizer = self.rust_tokenizer
302
        self.assertEqual(pyth_tokenizer.encode(""), [2])
303
        self.assertEqual(rust_tokenizer.encode(""), [2])
304

305
        self.assertEqual(pyth_tokenizer.encode("<s>"), [2, 204])
306
        self.assertEqual(rust_tokenizer.encode("<s>"), [2, 204])
307

308
    @unittest.skipIf(
309
        os.getenv("RUN_TOKENIZER_INTEGRATION", "0") == "0",
310
        "RUN_TOKENIZER_INTEGRATION=1 to run tokenizer integration tests",
311
    )
312
    def test_integration_test_xnli(self):
313
        import tqdm
314

315
        pyth_tokenizer = self.tokenizer
316
        rust_tokenizer = self.rust_tokenizer
317

318
        dataset = load_dataset("code_x_glue_ct_code_to_text", "go")
319
        for item in tqdm.tqdm(dataset["validation"]):
320
            string = item["code"]
321
            encoded1 = pyth_tokenizer.encode(string)
322
            encoded2 = rust_tokenizer.encode(string)
323

324
            self.assertEqual(encoded1, encoded2)
325

326
            decoded1 = pyth_tokenizer.decode(encoded1, skip_special_tokens=True)
327
            decoded2 = rust_tokenizer.decode(encoded1, skip_special_tokens=True)
328

329
            self.assertEqual(decoded1, decoded2)
330

331
        dataset = load_dataset("xnli", "all_languages")
332

333
        for item in tqdm.tqdm(dataset["train"]):
334
            for string in item["premise"].values():
335
                encoded1 = pyth_tokenizer.encode(string)
336
                encoded2 = rust_tokenizer.encode(string)
337

338
                self.assertEqual(encoded1, encoded2)
339

340
                decoded1 = pyth_tokenizer.decode(encoded1, skip_special_tokens=True)
341
                decoded2 = rust_tokenizer.decode(encoded2, skip_special_tokens=True)
342

343
                self.assertEqual(decoded1, decoded2)
344

345
    def test_special_token_special_word(self):
346
        # the word inform should be split as ['in', 'form']
347
        tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
348
        tokenizer.add_tokens([AddedToken("<REPR_END>", rstrip=True, lstrip=True)], special_tokens=False)
349
        out1 = tokenizer.decode(
350
            tokenizer.encode("<REPR_END>inform", add_special_tokens=False), spaces_between_special_tokens=False
351
        )
352
        self.assertEqual(out1, "<REPR_END>inform")
353
        out2 = tokenizer.decode(
354
            tokenizer.encode("<REPR_END>inform", add_special_tokens=False), spaces_between_special_tokens=True
355
        )
356
        # decoding strips the added prefix space.
357
        self.assertEqual(out2, "<REPR_END> inform")
358
        input_ids = tokenizer.encode("<REPR_END>inform", add_special_tokens=False)
359
        self.assertEqual(input_ids, [256000, 43910])
360

361
        out2 = tokenizer.decode(
362
            tokenizer.encode(" <REPR_END> inform", add_special_tokens=False), spaces_between_special_tokens=False
363
        )
364
        # TODO @ArthurZ currently we strip left and right, so this will not keep the spaces
365
        self.assertEqual(out2, "<REPR_END>inform")
366

367
        ### Let's make sure decoding does not add extra spaces here and there
368
        # TODO @ArthurZ this should be affected by the lstrip/rstrip/single word /normalize refactoring
369
        # Since currently we always strip left and right of the token, results are as such
370
        input_ids = tokenizer.encode("<s> Hello<s>how", add_special_tokens=False)
371
        self.assertEqual(input_ids, [204, 25957, 204, 1139])
372
        tokens = tokenizer.tokenize("<s> Hello<s>how", add_special_tokens=False)
373
        self.assertEqual(tokens, ["<s>", "▁Hello", "<s>", "how"])
374
        decoded_tokens = tokenizer.decode(input_ids)
375
        self.assertEqual(decoded_tokens, "<s> Hello<s>how")
376

377
        # Let's make sure that if there are any spaces, we don't remove them!
378
        input_ids = tokenizer.encode(" <s> Hello<s> how", add_special_tokens=False)
379
        self.assertEqual(input_ids, [235248, 204, 25957, 204, 1368])
380
        tokens = tokenizer.tokenize(" <s> Hello<s> how", add_special_tokens=False)
381
        self.assertEqual(tokens, ["▁", "<s>", "▁Hello", "<s>", "▁how"])
382
        decoded_tokens = tokenizer.decode(input_ids)
383
        self.assertEqual(decoded_tokens, " <s> Hello<s> how")
384

385
    def test_some_edge_cases(self):
386
        tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
387

388
        sp_tokens = tokenizer.sp_model.encode("<s>>", out_type=str)
389
        self.assertEqual(sp_tokens, ["<s>", ">"])
390
        tokens = tokenizer.tokenize("<s>>")
391
        self.assertEqual(sp_tokens, tokens)
392
        self.assertEqual(tokens, ["<s>", ">"])
393

394
        tokens = tokenizer.tokenize("")
395
        self.assertEqual(tokens, [])
396
        self.assertEqual(tokens, tokenizer.sp_model.encode("", out_type=str))
397

398
        tokens = tokenizer.tokenize(" ")
399
        self.assertEqual(tokens, ["▁"])
400
        # a dummy prefix space is not added by the sp_model as it was de-activated
401
        self.assertEqual(tokens, tokenizer.sp_model.encode(" ", out_type=str))
402

403
        tokens = tokenizer.tokenize("▁")
404
        self.assertEqual(tokens, ["▁"])
405
        # a dummy prefix space is not added by the sp_model as it was de-activated
406
        self.assertEqual(tokens, tokenizer.sp_model.encode("▁", out_type=str))
407

408
        tokens = tokenizer.tokenize(" ▁")
409
        self.assertEqual(tokens, ["▁▁"])
410
        # a dummy prefix space is not added by the sp_model as it was de-activated
411
        self.assertEqual(tokens, tokenizer.sp_model.encode("▁▁", out_type=str))
412

413
    @require_jinja
414
    def test_tokenization_for_chat(self):
415
        tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
416

417
        test_chats = [
418
            [{"role": "user", "content": "Hello!"}],
419
            [
420
                {"role": "user", "content": "Hello!"},
421
                {"role": "assistant", "content": "Nice to meet you."},
422
            ],
423
            [{"role": "user", "content": "Hello!"}],
424
        ]
425
        # Matt: The third test case tests the default system message, but if this is ever changed in the
426
        #       class/repo code then that test will fail, and the case will need to be updated.
427
        tokenized_chats = [tokenizer.apply_chat_template(test_chat) for test_chat in test_chats]
428
        expected_tokens = [[235322, 235371, 571, 235298, 2997, 73786, 1645, 108, 4521, 149907, 235371, 571, 235298, 615, 73786, 108], [235322, 235371, 571, 235298, 2997, 73786, 1645, 108, 4521, 149907, 235371, 571, 235298, 615, 73786, 108, 235322, 235371, 571, 235298, 2997, 73786, 105776, 108, 7731, 577, 4664, 692, 35606, 235371, 571, 235298, 615, 73786, 108], [235322, 235371, 571, 235298, 2997, 73786, 1645, 108, 4521, 149907, 235371, 571, 235298, 615, 73786, 108]]  # fmt: skip
429
        for tokenized_chat, expected_tokens in zip(tokenized_chats, expected_tokens):
430
            self.assertListEqual(tokenized_chat, expected_tokens)
431

432

433
@require_sentencepiece
434
@require_tokenizers
435
class CommonSpmIntegrationTests(unittest.TestCase):
436
    """
437
    A class that regroups important test to make sure that we properly handle the special tokens.
438
    """
439

440
    def test_edge_case_tabulation(self):
441
        fast_tokenizer = GemmaTokenizerFast.from_pretrained("hf-internal-testing/dummy-gemma")
442
        slow_tokenizer = GemmaTokenizer.from_pretrained("hf-internal-testing/dummy-gemma")
443
        input_text = "Hey<eos>. \t\t \n\nyou  é  @#😈  🤗!       , 1234 15 5,61"
444
        EXPECTED_IDS = [ 2, 6750, 1, 235265, 235248, 255969, 235248, 109, 4747, 139, 235335, 139, 216311, 241316, 139, 239880, 235341, 144, 235269, 235248, 235274, 235284, 235304, 235310, 235248, 235274, 235308, 235248, 235308, 235269, 235318, 235274]  # fmt: skip
445
        EXPECTED_TOKENS = [ "Hey", "<eos>", ".", "▁", "\t\t", "▁", "\n\n", "you", "▁▁", "é", "▁▁", "@#", "😈", "▁▁", "🤗", "!", "▁▁▁▁▁▁▁", ",", "▁", "1", "2", "3", "4", "▁", "1", "5", "▁", "5", ",", "6", "1"]  # fmt: skip
446

447
        tokens = fast_tokenizer.tokenize(input_text)
448
        with self.subTest("test fast edge case fast"):
449
            self.assertEqual(tokens, EXPECTED_TOKENS)
450

451
        tokens = slow_tokenizer.tokenize(input_text)
452
        with self.subTest("test fast edge case fast"):
453
            self.assertEqual(tokens, EXPECTED_TOKENS)
454

455
        input_ids = fast_tokenizer.encode(input_text)
456
        with self.subTest("test fast edge case fast"):
457
            self.assertEqual(input_ids, EXPECTED_IDS)
458

459
        input_ids = slow_tokenizer.encode(input_text)
460
        with self.subTest("test fast edge case fast"):
461
            self.assertEqual(input_ids, EXPECTED_IDS)
462

463
        text = fast_tokenizer.decode(EXPECTED_IDS)
464
        with self.subTest("test fast edge case fast"):
465
            self.assertEqual(text, "<bos>Hey<eos>. \t\t \n\nyou  é  @#😈  🤗!       , 1234 15 5,61")
466

467
        text = slow_tokenizer.decode(EXPECTED_IDS)
468
        with self.subTest("test fast edge case fast"):
469
            self.assertEqual(text, "<bos>Hey<eos>. \t\t \n\nyou  é  @#😈  🤗!       , 1234 15 5,61")
470

471
        input_text = "\t\t\t\t \n\n61"
472
        EXPECTED_IDS = [2, 255971, 235248, 109, 235318, 235274]
473
        EXPECTED_TOKENS = ["\t\t\t\t", "▁", "\n\n", "6", "1"]
474

475
        tokens = fast_tokenizer.tokenize(input_text)
476
        with self.subTest("test fast edge case fast"):
477
            self.assertEqual(tokens, EXPECTED_TOKENS)
478

479
        tokens = slow_tokenizer.tokenize(input_text)
480
        with self.subTest("test fast edge case fast"):
481
            self.assertEqual(tokens, EXPECTED_TOKENS)
482

483
        input_ids = fast_tokenizer.encode(input_text)
484
        with self.subTest("test fast edge case fast"):
485
            self.assertEqual(input_ids, EXPECTED_IDS)
486

487
        input_ids = slow_tokenizer.encode(input_text)
488
        with self.subTest("test fast edge case fast"):
489
            self.assertEqual(input_ids, EXPECTED_IDS)
490

491
        text = fast_tokenizer.decode(EXPECTED_IDS)
492
        with self.subTest("test fast edge case fast"):
493
            self.assertEqual(text, "<bos>\t\t\t\t \n\n61")
494

495
        text = slow_tokenizer.decode(EXPECTED_IDS)
496
        with self.subTest("test fast edge case fast"):
497
            self.assertEqual(text, "<bos>\t\t\t\t \n\n61")
498