[example/glue] fix compute_metrics_fn for bart like models (#7248)

* fix compute_metrics_fn

* p.predictions -> preds

* apply suggestions

examples/text-classification/run_glue.py

@@ -150,10 +150,11 @@ def main():
 
     def build_compute_metrics_fn(task_name: str) -> Callable[[EvalPrediction], Dict]:
         def compute_metrics_fn(p: EvalPrediction):
+            preds = p.predictions[0] if type(p.predictions) == tuple else p.predictions
             if output_mode == "classification":
-                preds = np.argmax(p.predictions, axis=1)
-            elif output_mode == "regression":
-                preds = np.squeeze(p.predictions)
+                preds = np.argmax(preds, axis=1)
+            else:  # regression
+                preds = np.squeeze(preds)
             return glue_compute_metrics(task_name, preds, p.label_ids)
 
         return compute_metrics_fn