Fix `FillMaskPipelineTests` (#22894)

* fix

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

tests/models/flaubert/test_modeling_flaubert.py

@@ -394,11 +394,7 @@ class FlaubertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
     def is_pipeline_test_to_skip(
         self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
     ):
-        if pipeline_test_casse_name == "FillMaskPipelineTests":
-            # Get `ValueError: AttributeError: 'NoneType' object has no attribute 'new_ones'` or `AssertionError`.
-            # `FlaubertConfig` was never used in pipeline tests: cannot create a simple tokenizer
-            return True
-        elif (
+        if (
             pipeline_test_casse_name == "QAPipelineTests"
             and tokenizer_name is not None
             and not tokenizer_name.endswith("Fast")

tests/models/flaubert/test_modeling_tf_flaubert.py

@@ -310,11 +310,7 @@ class TFFlaubertModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.Test
     def is_pipeline_test_to_skip(
         self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
     ):
-        if pipeline_test_casse_name == "FillMaskPipelineTests":
-            # Get `ValueError: AttributeError: 'NoneType' object has no attribute 'new_ones'` or `AssertionError`.
-            # `FlaubertConfig` was never used in pipeline tests: cannot create a simple tokenizer
-            return True
-        elif (
+        if (
             pipeline_test_casse_name == "QAPipelineTests"
             and tokenizer_name is not None
             and not tokenizer_name.endswith("Fast")

tests/models/xlm/test_modeling_tf_xlm.py

@@ -313,11 +313,7 @@ class TFXLMModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase)
     def is_pipeline_test_to_skip(
         self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
     ):
-        if pipeline_test_casse_name == "FillMaskPipelineTests":
-            # Get `ValueError: AttributeError: 'NoneType' object has no attribute 'new_ones'` or `AssertionError`.
-            # `XLMConfig` was never used in pipeline tests: cannot create a simple tokenizer
-            return True
-        elif (
+        if (
             pipeline_test_casse_name == "QAPipelineTests"
             and tokenizer_name is not None
             and not tokenizer_name.endswith("Fast")

tests/models/xlm/test_modeling_xlm.py

@@ -395,11 +395,7 @@ class XLMModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin,
     def is_pipeline_test_to_skip(
         self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
     ):
-        if pipeline_test_casse_name == "FillMaskPipelineTests":
-            # Get `ValueError: AttributeError: 'NoneType' object has no attribute 'new_ones'` or `AssertionError`.
-            # `XLMConfig` was never used in pipeline tests: cannot create a simple tokenizer
-            return True
-        elif (
+        if (
             pipeline_test_casse_name == "QAPipelineTests"
             and tokenizer_name is not None
             and not tokenizer_name.endswith("Fast")

tests/pipelines/test_pipelines_fill_mask.py

@@ -302,7 +302,8 @@ class FillMaskPipelineTests(unittest.TestCase):
         )
         target_ids = {vocab[el] for el in targets}
         self.assertEqual({el["token"] for el in outputs}, target_ids)
-        self.assertEqual({el["token_str"] for el in outputs}, set(targets))
+        processed_targets = [tokenizer.decode([x]) for x in target_ids]
+        self.assertEqual({el["token_str"] for el in outputs}, set(processed_targets))
 
         # Call argument
         fill_masker = FillMaskPipeline(model=model, tokenizer=tokenizer)
@@ -316,22 +317,27 @@ class FillMaskPipelineTests(unittest.TestCase):
         )
         target_ids = {vocab[el] for el in targets}
         self.assertEqual({el["token"] for el in outputs}, target_ids)
-        self.assertEqual({el["token_str"] for el in outputs}, set(targets))
+        processed_targets = [tokenizer.decode([x]) for x in target_ids]
+        self.assertEqual({el["token_str"] for el in outputs}, set(processed_targets))
 
         # Score equivalence
         outputs = fill_masker(f"This is a {tokenizer.mask_token}", targets=targets)
         tokens = [top_mask["token_str"] for top_mask in outputs]
         scores = [top_mask["score"] for top_mask in outputs]
 
+        # For some BPE tokenizers, `</w>` is removed during decoding, so `token_str` won't be the same as in `targets`.
+        if set(tokens) == set(targets):
             unmasked_targets = fill_masker(f"This is a {tokenizer.mask_token}", targets=tokens)
             target_scores = [top_mask["score"] for top_mask in unmasked_targets]
             self.assertEqual(nested_simplify(scores), nested_simplify(target_scores))
 
         # Raises with invalid
-        with self.assertRaises(ValueError):
-            outputs = fill_masker(f"This is a {tokenizer.mask_token}", targets=[""])
         with self.assertRaises(ValueError):
             outputs = fill_masker(f"This is a {tokenizer.mask_token}", targets=[])
+        # For some tokenizers, `""` is actually in the vocabulary and the expected error won't raised
+        if "" not in tokenizer.get_vocab():
+            with self.assertRaises(ValueError):
+                outputs = fill_masker(f"This is a {tokenizer.mask_token}", targets=[""])
             with self.assertRaises(ValueError):
                 outputs = fill_masker(f"This is a {tokenizer.mask_token}", targets="")
 
@@ -368,8 +374,9 @@ class FillMaskPipelineTests(unittest.TestCase):
         # If we use the most probably targets, and filter differently, we should still
         # have the same results
         targets2 = [el["token_str"] for el in sorted(outputs, key=lambda x: x["score"], reverse=True)]
+        # For some BPE tokenizers, `</w>` is removed during decoding, so `token_str` won't be the same as in `targets`.
+        if set(targets2).issubset(targets):
             outputs2 = fill_masker(f"This is a {tokenizer.mask_token}", top_k=3, targets=targets2)
-
             # They should yield exactly the same result
             self.assertEqual(nested_simplify(outputs), nested_simplify(outputs2))