Workaround a known issue with control dependency on external tensors

This fix tensorflow_models/official/nlp/bert/run_pretraining failure with
explicit allreduce.

PiperOrigin-RevId: 338521478

Author: crccw

official/nlp/optimization.py

@@ -199,7 +199,7 @@ def _resource_apply_dense(self, grad, var, apply_state=None):
     # backward pass.
     # TODO(b/171088214): Remove it after the control dependency in
     # nested function is fixed.
-    with tf.control_dependencies([grad]):
+    with tf.control_dependencies([tf.identity(grad)]):
       lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
       decay = self._decay_weights_op(var, lr_t, apply_state)
     with tf.control_dependencies([decay]):
@@ -212,7 +212,7 @@ def _resource_apply_sparse(self, grad, var, indices, apply_state=None):
     # backward pass.
     # TODO(b/171088214): Remove it after the control dependency in
     # nested function is fixed.
-    with tf.control_dependencies([grad]):
+    with tf.control_dependencies([tf.identity(grad)]):
       lr_t, kwargs = self._get_lr(var.device, var.dtype.base_dtype, apply_state)
       decay = self._decay_weights_op(var, lr_t, apply_state)
     with tf.control_dependencies([decay]):