TensorHandlePtr combined_value = CreatePerDeviceValues(
      context.get(), in_components, device_name, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Loop to make synchronization failures more deterministic
  for (int i = 0; i < 100; ++i) {
    TensorHandlePtr multiply_result(
        Multiply(context.get(), combined_value.get(), combined_value.get(),
                 status.get()));

  TFE_TensorHandle* t = TFE_TensorHandleCopyToDevice(
      tensor, context, dev->underlying_device.c_str(), status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  auto dst = std::make_unique<LoggedTensor>(t);
  *(dev->arrived_flag) = true;
  return MakeLoggedTensorHandle(context, dev->device_name, std::move(dst),
                                status);
}

  int num_retvals = 1;
  std::vector<AbstractTensorHandle*> outputs(1);
  GradientRegistry registry;
  s = RegisterGradients(&registry);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto tape = std::make_unique<Tape>(/*persistent=*/false);
  s = Execute(check_numerics_op.get(), ctx.get(), absl::MakeSpan(outputs),
              &num_retvals, &forward_op, tape.get(), registry);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

    (*var_to_shape_map)[key] = TensorShape(entry.shape());
    (*var_to_data_type_map)[key] = DataType(entry.dtype());
  }
  // The returned pointers are owned by the caller.
  return std::make_pair(std::move(var_to_shape_map),
                        std::move(var_to_data_type_map));
}

}  // namespace checkpoint