diff --git a/Vision/classification/image/resnet50/examples/train_eager_consistent.sh b/Vision/classification/image/resnet50/examples/train_eager_consistent.sh
index 2b6aac5fe..3245fb67c 100644
--- a/Vision/classification/image/resnet50/examples/train_eager_consistent.sh
+++ b/Vision/classification/image/resnet50/examples/train_eager_consistent.sh
@@ -45,3 +45,4 @@ python3 -m oneflow.distributed.launch \
         --num-epochs $EPOCH \
         --train-batch-size $TRAIN_BATCH_SIZE \
         --val-batch-size $VAL_BATCH_SIZE \
+        --scale-grad
diff --git a/Vision/classification/image/resnet50/train.py b/Vision/classification/image/resnet50/train.py
index c1ba49ba4..459a4b421 100644
--- a/Vision/classification/image/resnet50/train.py
+++ b/Vision/classification/image/resnet50/train.py
@@ -257,6 +257,7 @@ def train_one_epoch(self):
             else:
                 top1_acc = 0
 
+            loss = loss * self.world_size
             self.meter_train_iter(loss, top1_acc)
 
             self.cur_batch += 1
@@ -269,7 +270,7 @@ def train_eager(self):
         if loss.is_global and self.scale_grad:
             # NOTE(zwx): scale init grad with world_size
             # because global_tensor.mean() include dividor numel * world_size
-            loss = loss / self.world_size
+            # loss = loss / self.world_size
             loss.backward()
             for param_group in self.optimizer.param_groups:
                 for param in param_group.parameters: