fix unittest test=develop

paddle/fluid/framework/parallel_executor.cc

@@ -197,9 +197,17 @@ ParallelExecutor::ParallelExecutor(
     PADDLE_ENFORCE(places.size() > 1,
                    "If you set build_strategy.reduce with 'Reduce',"
                    "the number of places must be greater than 1.");
-    PADDLE_ENFORCE(exec_strategy.type_ != ExecutionStrategy::kParallelGraph,
-                   "You should set build_strategy.reduce with 'AllReduce' for "
-                   "the ParallelGraph executor type");
+  }
+
+  if (exec_strategy.type_ == ExecutionStrategy::kParallelGraph) {
+    PADDLE_ENFORCE(
+        member_->use_all_reduce_,
+        "build_strategy.reduce should be `AllReduce` if you want to use"
+        "ParallelGraph executor.");
+    PADDLE_ENFORCE(
+        member_->use_cuda_,
+        "execution_strategy.use_cuda should be True if you want to use"
+        "ParallelGraph executor.");
   }
 
   // Step 1. Bcast the params to devs.

python/paddle/fluid/tests/unittests/test_parallel_executor_mnist.py

@@ -166,6 +166,8 @@ class TestMNIST(TestParallelExecutorBase):
     def check_batchnorm_fc_convergence(self, use_cuda, exec_type):
         if use_cuda and not core.is_compiled_with_cuda():
             return
+        if not use_cuda and exec_type == ExecutorType.ParallelGraph:
+            return
 
         img, label = self._init_data()
 

python/paddle/fluid/tests/unittests/test_parallel_executor_transformer.py

@@ -173,10 +173,6 @@ class TestTransformer(TestParallelExecutorBase):
     def test_main(self):
         if core.is_compiled_with_cuda():
             self.check_network_convergence(transformer, use_cuda=True)
-            self.check_network_convergence(
-                transformer,
-                use_cuda=True,
-                exec_type=ExecutorType.ParallelGraph)
             self.check_network_convergence(
                 transformer, use_cuda=True, enable_sequential_execution=True)
         self.check_network_convergence(transformer, use_cuda=False, iter=5)