diff --git a/src/transformers/pipelines/base.py b/src/transformers/pipelines/base.py
index 124f2e290ebc..2139980da236 100644
--- a/src/transformers/pipelines/base.py
+++ b/src/transformers/pipelines/base.py
@@ -590,6 +590,7 @@ def _parse_and_tokenize(
             return_tensors=self.framework,
             padding=padding,
             truncation=truncation,
+            **kwargs,
         )
 
         return inputs
diff --git a/tests/test_pipelines_fill_mask.py b/tests/test_pipelines_fill_mask.py
index f087ed213577..fec273f484b3 100644
--- a/tests/test_pipelines_fill_mask.py
+++ b/tests/test_pipelines_fill_mask.py
@@ -77,9 +77,12 @@ def test_torch_fill_mask(self):
         self.assertIsInstance(outputs, list)
 
         # This used to fail with `cannot mix args and kwargs`
-        outputs = nlp(valid_inputs, something=False)
+        outputs = nlp(valid_inputs, verbose=False)
         self.assertIsInstance(outputs, list)
 
+        # This fails because `invalid_arg` is not an argument of the pipeline's tokenizer
+        self.assertRaises(TypeError, nlp, valid_inputs, invalid_arg=False)
+
     @require_torch
     def test_torch_fill_mask_with_targets(self):
         valid_inputs = ["My name is <mask>"]