fix dist transpiler bug

python/paddle/fluid/distribute_transpiler.py

@@ -278,6 +278,15 @@ class DistributeTranspiler:
             # we don't need to create them when grad arrives.
             # change client side var name to origin name by
             # removing ".trainer_%d" suffix
+            # NOTE: single_trainer_var must be created for multi-trainer
+            # case to merge grads from multiple trainers
+            single_trainer_var = \
+                pserver_program.global_block().create_var(
+                    name=orig_var_name,
+                    persistable=True,
+                    type=v.type,
+                    dtype=v.dtype,
+                    shape=v.shape)
             suff_idx = v.name.find(".trainer_")
             if suff_idx >= 0:
                 orig_var_name = v.name[:suff_idx]
@@ -293,12 +302,6 @@ class DistributeTranspiler:
                         shape=v.shape)
                     recv_inputs.append(var)
             else:
-                single_trainer_var = pserver_program.global_block().create_var(
-                    name=orig_var_name,
-                    persistable=True,
-                    type=v.type,
-                    dtype=v.dtype,
-                    shape=v.shape)
                 recv_inputs.append(single_trainer_var)
 
         # step3