Fixing question-answering with long contexts (#13873)

* Tmp.

* Fixing BC for question answering with long context.

* Capping model_max_length to avoid tf overflow.

* Bad workaround bugged roberta.

* Fixing name.

tests/test_modeling_led.py

@@ -162,6 +162,11 @@ class LEDModelTester:
             attention_window=self.attention_window,
         )
 
+    def get_pipeline_config(self):
+        config = self.get_config()
+        config.max_position_embeddings = 100
+        return config
+
     def prepare_config_and_inputs_for_common(self):
         config, inputs_dict = self.prepare_config_and_inputs()
         global_attention_mask = torch.zeros_like(inputs_dict["input_ids"])