std::vector<TFE_TensorHandle*> device_inputs;
    device_inputs.reserve(inputs.size());
    for (int input_index = 0; input_index < inputs.size(); ++input_index) {
      // Parallel tensors are divided between operations by device.
      device_inputs.push_back(inputs[input_index]->tensor(device_index));
    }
    device_thread->StartExecute(
        context, operation_name, std::move(device_inputs), attributes,

/// this file to ensure a compliant plugin can be built and that the plugin can
/// be used without recompilation in the widest range of TensorFlow versions.
///
/// The header is divided into sections, as follows:
///   1. Opaque plugin private data structures and wrappers for type safety;
///   2. Function tables for plugin functionality;
///   3. Versioning metadata;

        passed to the optimizer but the reported loss will be a scalar value.
    *   `SUM`: Scalar sum of weighted losses. 4. `SUM_OVER_BATCH_SIZE`: Scalar
        `SUM` divided by number of elements in losses. This reduction type is
        not supported when used with `tf.distribute.Strategy` outside of
        built-in training loops like `tf.keras` `compile`/`fit`.