fix bug for fused_linear_grad_add and main_grad (#56030)

python/paddle/distributed/fleet/utils/mix_precision_utils.py

@@ -55,6 +55,8 @@ class MixPrecisionLayer(nn.Layer):
             ), "In main_grad node, param.grad should be None, but find param[{}] has grad.".format(
                 param.name
             )
+            if tmp_grad._is_initialized():
+                # Some previous pylayer may return None, should check grad validation.
                 if param.main_grad is None:
                     param.main_grad = core.eager.Tensor(
                         value=tmp_grad.cast(paddle.float32).value(),