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_conformer/test_modeling_wav2vec2_conformer.py

@@ -23,6 +23,7 @@ from datasets import load_dataset
 
 from transformers import Wav2Vec2ConformerConfig, is_torch_available
 from transformers.testing_utils import (
+    is_flaky,
     is_pt_flax_cross_test,
     require_torch,
     require_torch_accelerator,
@@ -452,6 +453,12 @@ class Wav2Vec2ConformerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest
         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 `Wav2Vec2ConformerGumbelVectorQuantizer.forward` is not stable."
+    )
+    def test_batching_equivalence(self):
+        super().test_batching_equivalence()
+
     def test_model_with_relative(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs(position_embeddings_type="relative")
         self.model_tester.create_and_check_model(*config_and_inputs)