diff --git a/returnn/tensor/_tensor_extra.py b/returnn/tensor/_tensor_extra.py
index da3751c690..fae81301db 100644
--- a/returnn/tensor/_tensor_extra.py
+++ b/returnn/tensor/_tensor_extra.py
@@ -17,8 +17,9 @@
 from .dim import Dim, batch_dim, VerifyOutShapeException
 import returnn.tensor.tensor as _t
 import returnn.tensor.marked_dim as _m
+
+# noinspection PyProtectedMember
 import returnn.frontend._api as _frontend_api
-import returnn._internal_frontend_api as _internal_frontend_api
 from ._tensor_mixin_base import _TensorMixinBase
 
 
@@ -214,9 +215,9 @@ def _handle_extra_kwargs(
             # even in graph-based backends.
             # Rather, the logic to create placeholders should be done elsewhere.
             # noinspection PyProtectedMember
-            from returnn.tf.frontend_low_level._internal_frontend import TFInternalFrontend
+            from returnn.tf.frontend_low_level._frontend import TFFrontend
 
-            self.raw_tensor = TFInternalFrontend.create_placeholder(self)
+            self.raw_tensor = TFFrontend.create_placeholder(self)
             # Do that after same_dim_tags_as handling.
             _auto_create_size_placeholders_on_dim_tags(name=self.name, dim_tags=self._dims)
 
@@ -233,13 +234,13 @@ def raw_frontend(self) -> Optional[Type[_frontend_api.Frontend]]:
         return _frontend_api.get_frontend_by_raw_tensor_type(type(self._raw_tensor))
 
     @property
-    def _raw_internal_frontend(self) -> Optional[Type[_internal_frontend_api.InternalFrontend]]:
+    def _raw_internal_frontend(self) -> Optional[Type[_frontend_api.Frontend]]:
         """
         :return: the internal backend for the raw tensor
         """
         if self._raw_tensor is None:
             return None
-        return _internal_frontend_api.get_internal_frontend_by_tensor_type(type(self._raw_tensor))
+        return _frontend_api.get_frontend_by_raw_tensor_type(type(self._raw_tensor))
 
     @property
     def control_flow_ctx(self) -> Optional[ControlFlowContext]:
@@ -2645,12 +2646,13 @@ def get_sequence_mask(self):
         :return: seq mask of shape (batch,time) if we are batch-major, else (time,batch) if we are time-major
         :rtype: tf.Tensor
         """
-        from returnn._internal_frontend_api import get_internal_frontend_by_tensor_type
+        # noinspection PyProtectedMember
+        from returnn.frontend._api import get_frontend_by_raw_tensor_type
 
         assert self.time_dim_axis is not None
         assert self.batch_dim_axis is not None
         dyn_seq_len = self.get_sequence_lengths()
-        rf = get_internal_frontend_by_tensor_type(type(dyn_seq_len))
+        rf = get_frontend_by_raw_tensor_type(type(dyn_seq_len))
 
         if self.is_time_major:
             assert self.batch_dim_axis == 1
diff --git a/returnn/tensor/_tensor_op_overloads.py b/returnn/tensor/_tensor_op_overloads.py
index 3807be0456..d30db2c14f 100644
--- a/returnn/tensor/_tensor_op_overloads.py
+++ b/returnn/tensor/_tensor_op_overloads.py
@@ -8,6 +8,7 @@
 if TYPE_CHECKING:
     from .tensor import Tensor  # just for type hints; otherwise use _t.Tensor
 
+# noinspection PyProtectedMember
 import returnn.frontend._api as _frontend_api
 from ._tensor_mixin_base import _TensorMixinBase