Compatible upgrade of sparse_momentum for master param (#107)

Author: GuoxiaWang

dynamic/utils/hybrid_optimizer.py

@@ -85,12 +85,20 @@ def _append_optimize_op(self, block, param_and_grad):
         if getattr(param_and_grad[0], 'is_sparse_grad', None):
             index = getattr(param_and_grad[0], 'index', None)
             axis = getattr(param_and_grad[0], 'axis', None)
-            _, _ = paddle._C_ops.sparse_momentum(
-                param_and_grad[0], param_and_grad[1], velocity_acc, index, lr,
-                param_and_grad[0], velocity_acc, 'mu', self._momentum,
-                'use_nesterov', self._use_nesterov, 'regularization_method',
-                self._regularization_method, 'regularization_coeff',
-                self._regularization_coeff, 'axis', axis)
+            try:
+                _, _ = paddle._C_ops.sparse_momentum(
+                    param_and_grad[0], param_and_grad[1], velocity_acc, index, lr,
+                    param_and_grad[0], velocity_acc, 'mu', self._momentum,
+                    'use_nesterov', self._use_nesterov, 'regularization_method',
+                    self._regularization_method, 'regularization_coeff',
+                    self._regularization_coeff, 'axis', axis)
+            except:
+                _, _, _ = paddle._C_ops.sparse_momentum(
+                    param_and_grad[0], param_and_grad[1], velocity_acc, index, lr, master_weight,
+                    param_and_grad[0], velocity_acc, master_weight, 'mu', self._momentum,
+                    'use_nesterov', self._use_nesterov, 'regularization_method',
+                    self._regularization_method, 'regularization_coeff',
+                    self._regularization_coeff, 'axis', axis, 'multi_precision', find_master)
         else:
             _, _, _ = paddle._C_ops.momentum(
                 param_and_grad[0], param_and_grad[1], velocity_acc, lr,