Byebye test_batching_equivalence's flakiness (#35729)

* fix

* fix

* skip

* better error message

---------

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

tests/models/wav2vec2/test_modeling_wav2vec2.py

@@ -30,6 +30,7 @@ from transformers import Wav2Vec2Config, is_torch_available
 from transformers.testing_utils import (
     CaptureLogger,
     cleanup,
+    is_flaky,
     is_pt_flax_cross_test,
     is_pyctcdecode_available,
     is_torchaudio_available,
@@ -863,6 +864,12 @@ class Wav2Vec2RobustModelTest(ModelTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_model(*config_and_inputs)
 
+    @is_flaky(
+        description="The `codevector_idx` computed with `argmax()` in `Wav2Vec2GumbelVectorQuantizer.forward` is not stable."
+    )
+    def test_batching_equivalence(self):
+        super().test_batching_equivalence()
+
     def test_model_with_adapter(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_model_with_adapter(*config_and_inputs)