1. fix stop_gradient, 2. reduce_sum keep_dim

examples/erniesage/models/base.py

@@ -191,12 +191,12 @@ def all_gather(X):
     for i in range(trainer_num):
         copy_X = X * 1
         copy_X =  L.collective._broadcast(copy_X, i, True)
-        copy_X.stop_gradients=True
+        copy_X.stop_gradient=True
         Xs.append(copy_X)
 
     if len(Xs) > 1:
         Xs=L.concat(Xs, 0)
-        Xs.stop_gradients=True
+        Xs.stop_gradient=True
     else:
         Xs = Xs[0]
     return Xs

examples/erniesage/models/erniesage_v2.py

@@ -27,7 +27,7 @@ class ErnieSageV2(BaseNet):
             src_position_ids = L.expand(src_position_ids, [src_batch, 1, 1]) # [B, slot_seqlen * num_b, 1]
             zero = L.fill_constant([1], dtype='int64', value=0)
             input_mask = L.cast(L.equal(src_ids, zero), "int32")  # assume pad id == 0 [B, slot_seqlen, 1]
-            src_pad_len = L.reduce_sum(input_mask, 1) # [B, 1, 1]
+            src_pad_len = L.reduce_sum(input_mask, 1, keep_dim=True) # [B, 1, 1]
 
             dst_position_ids = L.reshape(
                 L.range(