diff --git a/torch_xla/tf_saved_model_integration.py b/torch_xla/tf_saved_model_integration.py
index 188653abc3f..511b9e02e9b 100644
--- a/torch_xla/tf_saved_model_integration.py
+++ b/torch_xla/tf_saved_model_integration.py
@@ -58,6 +58,19 @@ def _make_input_signatures(
         shape=spec.shape, dtype=getattr(tf, spec.dtype), name=f'args_{i}')
 
 
+def _mangle_tf_root_scope_name(name):
+  # TF has more restricted constrain on the variable names at root scope.
+  # Root scope name constrain: [A-Za-z0-9.][A-Za-z0-9_.\\-/]*
+  # Non-root scope name constrain: [A-Za-z0-9_.\\-/]*
+  # https://github.com/tensorflow/tensorflow/blob/51b601fa6bb7e801c0b6ae73c25580e40a8b5745/tensorflow/python/framework/ops.py#L3301-L3302
+  # The state_dict key doesn't have such constrain,
+  # the name need to be mangled when a root-scoped TF variable is created.
+  if name[0] in "._\\-/":
+    return 'k' + name
+  else:
+    return name
+
+
 def save_stablehlo_graph_as_tf(
     stablehlo_program: stablehlo.StableHLOGraphModule,
     path: os.PathLike,
@@ -81,7 +94,7 @@ def save_stablehlo_graph_as_tf(
   bundle = copy.deepcopy(stablehlo_program._bundle)
   tfm = tf.Module()
   bundle.state_dict = {
-      k: tf.Variable(v, trainable=False, name=k)
+      k: tf.Variable(v, trainable=False, name=_mangle_tf_root_scope_name(k))
       for k, v in bundle.state_dict.items()
   }
   bundle.additional_constants = [