transformers/tests/utils/test_convert_slow_tokenizer.py

import unittest
import warnings
from dataclasses import dataclass

from transformers.convert_slow_tokenizer import SpmConverter
from transformers.testing_utils import get_tests_dir


@dataclass
class FakeOriginalTokenizer:
    vocab_file: str


class ConvertSlowTokenizerTest(unittest.TestCase):
    def test_spm_converter_bytefallback_warning(self):
        spm_model_file_without_bytefallback = get_tests_dir("fixtures/test_sentencepiece.model")
        spm_model_file_with_bytefallback = get_tests_dir("fixtures/test_sentencepiece_with_bytefallback.model")

        original_tokenizer_without_bytefallback = FakeOriginalTokenizer(vocab_file=spm_model_file_without_bytefallback)

        with warnings.catch_warnings(record=True) as w:
            _ = SpmConverter(original_tokenizer_without_bytefallback)
        self.assertEqual(len(w), 0)

        original_tokenizer_with_bytefallback = FakeOriginalTokenizer(vocab_file=spm_model_file_with_bytefallback)

        with warnings.catch_warnings(record=True) as w:
            _ = SpmConverter(original_tokenizer_with_bytefallback)
        self.assertEqual(len(w), 1)

        self.assertIn(
            "The sentencepiece tokenizer that you are converting to a fast tokenizer uses the byte fallback option"
            " which is not implemented in the fast tokenizers.",
            str(w[0].message),
        )