add custom op gelu for bert amp training (#5008)

* add custom op gelu

* refine run_pretrain

PaddleNLP/benchmark/bert/run_pretrain.py

@@ -257,7 +257,7 @@ def do_train(args):
         ])
     if args.use_amp:
         amp_list = paddle.fluid.contrib.mixed_precision.AutoMixedPrecisionLists(
-            custom_white_list=['softmax'])
+            custom_white_list=['softmax', 'layer_norm', 'gelu'])
         optimizer = paddle.fluid.contrib.mixed_precision.decorate(
             optimizer,
             amp_list,

PaddleNLP/benchmark/bert/run_pretrain_single.py

@@ -228,7 +228,7 @@ def do_train(args):
         ])
     if args.use_amp:
         amp_list = paddle.fluid.contrib.mixed_precision.AutoMixedPrecisionLists(
-            custom_white_list=['layer_norm', 'softmax'])
+            custom_white_list=['layer_norm', 'softmax', 'gelu'])
         optimizer = paddle.fluid.contrib.mixed_precision.decorate(
             optimizer,
             amp_list,