keep the sub of in grad of embeddinglookup on device

mindspore/ops/_grad/grad_array_ops.py

@@ -194,23 +194,26 @@ def get_bprop_tile(self):
 @bprop_getters.register(inner.EmbeddingLookup)
 def get_bprop_embedding_lookup(self):
     """Generate bprop for EmbeddingLookup"""
-    host_sub = P.Sub().add_prim_attr('primitive_target', 'CPU')
+    sub_op = P.Sub()
+    reshape_op = P.Reshape()
     host_reshape = P.Reshape().add_prim_attr('primitive_target', 'CPU')
     def bprop_sparse(x, indices, offset, reduce_scatter_flag, split_num, out, dout):
         x_shp = shape_op(x)
-        if reduce_scatter_flag is True:
-            elu_grad = G.EmbeddingLookupCommGrad()
-            actual_dout = elu_grad(dout, split_num)
-        else:
-            actual_dout = dout
-        new_indices = host_sub(indices, offset)
+        new_indices = sub_op(indices, offset)
         # Reshape the 'new_indices'
         new_indices_shape_changed = (size_op(new_indices),)
-        new_indices = host_reshape(new_indices, new_indices_shape_changed)
-        # Reshape the 'actual_dout'
+        new_indices = reshape_op(new_indices, new_indices_shape_changed)
         x_shp_tail = x_shp[1:]
         actual_dout_shape_changed = new_indices_shape_changed + x_shp_tail
-        actual_dout = host_reshape(actual_dout, actual_dout_shape_changed)
+        if reduce_scatter_flag is True:
+            # On host
+            elu_grad = G.EmbeddingLookupCommGrad()
+            actual_dout = elu_grad(dout, split_num)
+            # Reshape the 'actual_dout' on host
+            actual_dout = host_reshape(actual_dout, actual_dout_shape_changed)
+        else:
+            # Reshape the 'actual_dout' on device
+            actual_dout = reshape_op(dout, actual_dout_shape_changed)
         return (new_indices, actual_dout, x_shp), zeros_like(indices), zeros_like(offset), \
                zeros_like(reduce_scatter_flag), zeros_like(split_num)
     return bprop_sparse