} else {
        // Generalize differing axis lengths to "variable"/"unknown".
        for (int axis_index = 0; axis_index < combined_shape.dims();
             ++axis_index) {
          int64_t axis_length = combined_shape.dim_size(axis_index);
          if (axis_length != component_shape.dim_size(axis_index)) {
            axis_length = -1;
          }
          TF_RETURN_IF_ERROR(

    // the purposes of running this graph in a session. Thus, we don't
    // record the source node as being modified.
    RecordMutation(graph, *dst.oper, "updating input tensor");
  }
}

// Apis that are corresponding to python c api. --------------------------

void TF_AddOperationControlInput(TF_Graph* graph, TF_Operation* op,
                                 TF_Operation* input) {

        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This
        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,

  TFE_TensorHandle* axis = TestAxisTensorHandle(ctx);
  TFE_Op* minOp = MinOp(ctx, input, axis);
  TFE_TensorHandle* retvals[1] = {nullptr};
  int num_retvals = 1;
  TFE_Execute(minOp, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteOp(minOp);
  TFE_DeleteTensorHandle(input);
  TFE_DeleteTensorHandle(axis);
  ASSERT_EQ(1, num_retvals);

      context.get(), components, device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  int num_axes = TFE_TensorHandleNumDims(combined_value.get(), status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  EXPECT_EQ(num_axes, 1);
}

TEST(PARALLEL_DEVICE, TestNestedParallelDevices) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(

struct TF_VariableInfo {
  TF_VariableInfo() = delete;
  // TF_VariableInfo is constructed here by TensorFlow, and will be passed to
  // plugin as a opaque pointer. Plugin will need to call C APIs below to
  // operate on TF_VariableInfo (such as allocate temp tensor for the `var` held
  // by the underlying tensorflow::VariableInfo.
  TF_VariableInfo(int index, const std::string& name, tensorflow::Var* var) {

                                        std::move(mixed_handles), status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK);
  // Can't take the shape of a parallel tensor with varying numbers of axes, but
  // running operations on them is OK.
  TF_ASSERT_OK(unknown_length_vector->Shape(&shape));
  EXPECT_THAT(*shape, ElementsAre(-1));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> size_op(

  // Make a placeholder operation.
  TF_Operation* feed = Placeholder(graph, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Make a constant operation with the scalar "0", for axis.
  TF_Operation* one = ScalarConst(0, graph, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Create a session for this graph.
  CSession csession(graph, s, use_XLA);

  return absl::OkStatus();
}

// Helper functions which delegate to `AbstractOperation`, update
// the state of the ForwardOperation and call the tape as appropriate.
// These APIs are mainly to facilitate testing and are subject to change.
namespace internal {
Status Reset(AbstractOperation* op_, const char* op,
             const char* raw_device_name, ForwardOperation* forward_op_) {

TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "Min", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, input, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpAddInput(op, axis, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);