diff --git a/tensorflow_recommenders_addons/dynamic_embedding/python/ops/dynamic_embedding_optimizer.py b/tensorflow_recommenders_addons/dynamic_embedding/python/ops/dynamic_embedding_optimizer.py
index eabf9103c..7018b4775 100644
--- a/tensorflow_recommenders_addons/dynamic_embedding/python/ops/dynamic_embedding_optimizer.py
+++ b/tensorflow_recommenders_addons/dynamic_embedding/python/ops/dynamic_embedding_optimizer.py
@@ -96,6 +96,9 @@ def apply_grad_to_update_var(var, grad):
         else:
           return update_op
       else:
+        if not var.params.trainable:
+          return control_flow_ops.no_op()
+
         with ops.colocate_with(None, ignore_existing=True):
           _slots = [self.get_slot(var, _s) for _s in self.get_slot_names()]
           var._track_optimizer_slots(_slots)
diff --git a/tensorflow_recommenders_addons/dynamic_embedding/python/ops/tf_patch.py b/tensorflow_recommenders_addons/dynamic_embedding/python/ops/tf_patch.py
index 5a482ce03..d31fc93b6 100644
--- a/tensorflow_recommenders_addons/dynamic_embedding/python/ops/tf_patch.py
+++ b/tensorflow_recommenders_addons/dynamic_embedding/python/ops/tf_patch.py
@@ -77,6 +77,9 @@ def update_op(self, optimizer, g):
     # pylint: disable=protected-access
     # for better convergence:
 
+    if not self._v.params.trainable:
+      return control_flow_ops.no_op()
+
     with ops.colocate_with(None, ignore_existing=True):
       _slots = [
           optimizer.get_slot(self._v, _s) for _s in optimizer.get_slot_names()