diff --git a/python/paddle/distributed/fleet/meta_optimizers/amp_optimizer.py b/python/paddle/distributed/fleet/meta_optimizers/amp_optimizer.py
index 24e0b196d4974ae9f8e3fe0612691aa53b48c2f3..c751e229cbbe2b900ead900297ff9956946b9e75 100644
--- a/python/paddle/distributed/fleet/meta_optimizers/amp_optimizer.py
+++ b/python/paddle/distributed/fleet/meta_optimizers/amp_optimizer.py
@@ -25,7 +25,6 @@ class AMPOptimizer(MetaOptimizerBase):
             "LarsOptimizer",
             "LambOptimizer",
             "RecomputeOptimizer",
-            "GradientMergeOptimizer",
             "GraphExecutionOptimizer",
         ]
         self.meta_optimizers_black_list = ["DGCOptimizer"]
diff --git a/python/paddle/distributed/fleet/meta_optimizers/gradient_merge_optimizer.py b/python/paddle/distributed/fleet/meta_optimizers/gradient_merge_optimizer.py
index 6315fbf5a0d633c576b033594f6ebc1e06228e31..380fbc2e09ebffe89cfaabfd8d753dc47e8d85ff 100644
--- a/python/paddle/distributed/fleet/meta_optimizers/gradient_merge_optimizer.py
+++ b/python/paddle/distributed/fleet/meta_optimizers/gradient_merge_optimizer.py
@@ -21,6 +21,7 @@ class GradientMergeOptimizer(MetaOptimizerBase):
         self.inner_opt = optimizer
         self.wrapped_opt = None
         self.meta_optimizers_white_list = [
+            "AMPOptimizer",
             "LarsOptimizer",
             "LambOptimizer",
             "GraphExecutionOptimizer",
diff --git a/python/paddle/fluid/contrib/mixed_precision/decorator.py b/python/paddle/fluid/contrib/mixed_precision/decorator.py
index 37996b6228efe4c2e211b027551bafb4c029d98a..2215d11aa06c2f81fdd1591a75a2e4acf23f2998 100644
--- a/python/paddle/fluid/contrib/mixed_precision/decorator.py
+++ b/python/paddle/fluid/contrib/mixed_precision/decorator.py
@@ -159,9 +159,6 @@ class OptimizerWithMixedPrecision(object):
             params_grads = self._optimizer.backward(
                 self._scaled_loss, startup_program, parameter_list, no_grad_set,
                 callbacks)
-            # Change the op_role_var attr for some ops, so that gradients
-            # transferred across GPUs can be FP16.
-            update_role_var_grad(train_program, params_grads)
         return params_grads
 
     def apply_gradients(self, params_grads):
@@ -176,6 +173,10 @@ class OptimizerWithMixedPrecision(object):
             A list of optimize operators.
         """
 
+        # Change the op_role_var attr for some ops, so that gradients
+        # transferred across GPUs can be FP16.
+        update_role_var_grad(self._train_program, params_grads)
+
         grads = [g for _, g in params_grads]
         if not self._is_distributed:
             with self._train_program._optimized_guard(grads):
diff --git a/python/paddle/fluid/tests/unittests/test_fleet_gradient_merge_meta_optimizer.py b/python/paddle/fluid/tests/unittests/test_fleet_gradient_merge_meta_optimizer.py
index a40bc9a9fba6efd60af34846426d33740212d286..2d03b267fe9e3c969a82976fd3309374c53efa46 100644
--- a/python/paddle/fluid/tests/unittests/test_fleet_gradient_merge_meta_optimizer.py
+++ b/python/paddle/fluid/tests/unittests/test_fleet_gradient_merge_meta_optimizer.py
@@ -46,6 +46,19 @@ class TestFleetGradientMergeMetaOptimizer(TestFleetMetaOptimizer):
         self.assertIn('@GradientMerge', ''.join(vars))
         self.assertIn('subprog', ''.join(vars))
 
+    def test_gm_amp_optimizer(self):
+        train_prog, startup_prog = paddle.fluid.Program(), paddle.fluid.Program(
+        )
+        avg_cost, strategy = self.net(train_prog, startup_prog)
+        self.set_strategy(strategy, 'gradient_merge')
+        self.set_strategy(strategy, 'amp')
+        self.optimizer(avg_cost, strategy, train_prog, startup_prog)
+        print(train_prog)
+
+        vars = [x.name for x in train_prog.list_vars()]
+        self.assertIn('@GradientMerge', ''.join(vars))
+        self.assertIn('cast', ''.join(vars))
+
 
 if __name__ == "__main__":
     unittest.main()