diff --git a/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/modeling.py b/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/modeling.py
index c32e2e67e4d99879ed5ec3c8aa33de04a75beeb2..74a83e90983d6d0286d716dbcf54e41dd42a5da1 100644
--- a/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/modeling.py
+++ b/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/modeling.py
@@ -31,7 +31,7 @@ from gpu_environment import get_custom_getter
 
 from npu_bridge.estimator.npu_unary_ops import npu_unary_ops
 from npu_bridge.estimator import npu_ops
-from npu_bridge.estimator.npu_aicore_ops import npu_aicore_ops
+#from npu_bridge.estimator.npu_aicore_ops import npu_aicore_ops
 
 class BertConfig(object):
   """Configuration for `BertModel`."""
@@ -289,6 +289,7 @@ def gelu(x):
 
   if tf.flags.FLAGS.npu_bert_fused_gelu:
       if tf.flags.FLAGS.use_fast_gelu:
+          from npu_bridge.estimator.npu_aicore_ops import npu_aicore_ops
           return npu_aicore_ops.fast_gelu(x)
       else:
           return npu_unary_ops.gelu(x)
@@ -388,6 +389,7 @@ def dropout(input_tensor, dropout_prob):
   if tf.flags.FLAGS.npu_bert_npu_dropout:
     output = npu_ops.dropout(input_tensor, 1.0 - dropout_prob)
   elif tf.flags.FLAGS.npu_bert_npu_dropout_v3:
+    from npu_bridge.estimator import npu_aicore_ops
     output = npu_aicore_ops.dropout_v3(input_tensor, 1.0 - dropout_prob)
   else:
     output = tf.nn.dropout(input_tensor, 1.0 - dropout_prob)
diff --git a/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/run_classifier.py b/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/run_classifier.py
index 113cecefc0ae67a15334c9f83938f3bc7625387f..8076c5db88c43a7e9bdf50da2e216efb71007757 100644
--- a/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/run_classifier.py
+++ b/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/run_classifier.py
@@ -95,9 +95,9 @@ flags.DEFINE_integer("npu_bert_loss_scale", 0,
 flags.DEFINE_bool("npu_bert_clip_by_global_norm", False,
                 "Use clip_by_global_norm if True, or use clip_by_norm for each gradient")
 
-flags.DEFINE_bool('npu_bert_npu_dropout', True, 'Whether to use npu defined dropout op')
+flags.DEFINE_bool('npu_bert_npu_dropout', False, 'Whether to use npu defined dropout op')
 
-flags.DEFINE_bool('npu_bert_npu_dropout_v3', False, 'Whether to use npu defined dropout_v3 op')
+flags.DEFINE_bool('npu_bert_npu_dropout_v3', True, 'Whether to use npu defined dropout_v3 op')
 
 flags.DEFINE_bool('npu_bert_tail_optimize', False, 'Whether to use npu allreduce tail optimization')
 
@@ -249,7 +249,7 @@ def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
       input_mask=input_mask,
       token_type_ids=segment_ids,
       use_one_hot_embeddings=use_one_hot_embeddings,
-      compute_type=tf.float32)
+      compute_type=tf.float16 if FLAGS.precision_mode == "allow_mix_precision" else tf.float32)
 
   # In the demo, we are doing a simple classification task on the entire
   # segment.
diff --git a/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/run_squad.py b/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/run_squad.py
index 56815803e955ee20039632a1b328797c951b1890..0d21f03d2242e77081ff04244ff7352504b9ca0e 100644
--- a/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/run_squad.py
+++ b/TensorFlow/built-in/nlp/BertNV_Series_for_TensorFlow/src/run_squad.py
@@ -194,9 +194,9 @@ def extract_run_squad_flags():
   flags.DEFINE_bool("npu_bert_clip_by_global_norm", False,
                     "Use clip_by_global_norm if True, or use clip_by_norm for each gradient")
 
-  flags.DEFINE_bool('npu_bert_npu_dropout', True, 'Whether to use npu defined dropout op')
+  flags.DEFINE_bool('npu_bert_npu_dropout', False, 'Whether to use npu defined dropout op')
 
-  flags.DEFINE_bool('npu_bert_npu_dropout_v3', False, 'Whether to use npu defined dropout_v3 op')
+  flags.DEFINE_bool('npu_bert_npu_dropout_v3', True, 'Whether to use npu defined dropout_v3 op')
 
   flags.DEFINE_bool('npu_bert_tail_optimize', False, 'Whether to use npu allreduce tail optimization')
 
@@ -241,7 +241,7 @@ def create_model(bert_config, is_training, input_ids, input_mask, segment_ids,
       input_mask=input_mask,
       token_type_ids=segment_ids,
       use_one_hot_embeddings=use_one_hot_embeddings,
-      compute_type=tf.float32)
+      compute_type=tf.float16 if FLAGS.precision_mode == "allow_mix_precision" else tf.float32)
 
   final_hidden = model.get_sequence_output()