Update message-efficient.rst (#2247)

docs/source/guide/message-efficient.rst

@@ -59,7 +59,7 @@ implementation would be like:
 
     linear = nn.Parameter(torch.FloatTensor(size=(1, node_feat_dim * 2)))
     def concat_message_function(edges):
-        {'cat_feat': torch.cat([edges.src.ndata['feat'], edges.dst.ndata['feat']])}
+         return {'cat_feat': torch.cat([edges.src.ndata['feat'], edges.dst.ndata['feat']])}
     g.apply_edges(concat_message_function)
     g.edata['out'] = g.edata['cat_feat'] * linear
 
@@ -87,4 +87,4 @@ The above two implementations are mathematically equivalent. The latter
 one is more efficient because it does not need to save feat_src and
 feat_dst on edges, which is not memory-efficient. Plus, addition could
 be optimized with DGL’s built-in function ``u_add_v``, which further
-speeds up computation and saves memory footprint.
\ No newline at end of file
+speeds up computation and saves memory footprint.