[AMP] Check call order of paddle.amp.decorate and paddle.DataParallel (…

…#38785)

* check amp.decorate and DataParallel

* refine coverage

* fix layer dtype

* refine code

python/paddle/fluid/dygraph/amp/auto_cast.py

@@ -145,6 +145,10 @@ def check_models(models):
             raise RuntimeError(
                 "Current train mode is pure fp16, models should be paddle.nn.Layer, but receive {}.".
                 format(type(model)))
+        if isinstance(model, paddle.DataParallel):
+            raise RuntimeError(
+                "For distributed AMP training, you should first use paddle.amp.decorate() to decotate origin model, and then call paddle.DataParallel get distributed model."
+            )
 
 
 def check_optimizers(optimizers):

python/paddle/fluid/dygraph/layers.py

@@ -1569,6 +1569,8 @@ def _apply(self, func, device, dtype, blocking, include_sublayers=True):
         for key, buf in self._buffers.items():
             self._buffers[key] = func(buf, device, dtype, blocking)
 
+        self._dtype = dtype
+
     def _to_impl(self,
                  device=None,
                  dtype=None,

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

@@ -536,6 +536,14 @@ def __init__(self):
 
         self.assertRaises(TypeError, test_error_model)
 
+        def test_error_distributed_model():
+            model = fluid.dygraph.Conv2D(3, 2, 3, bias_attr=False, act=None)
+            model = paddle.DataParallel(model)
+            with fluid.dygraph.guard():
+                model = paddle.amp.decorate(models=model, level='O2')
+
+        self.assertRaises(RuntimeError, test_error_distributed_model)
+
         def test_error_optimizer():
             class MyOptimizer(object):
                 def __init__(self):