ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
    TFE_DeleteContextOptions(opts);
    TFE_Executor* executor = TFE_NewExecutor(
        /*is_async=*/false, /*enable_streaming_enqueue=*/true,
        /*in_flight_nodes_limit=*/0);
    TFE_ContextSetExecutorForThread(ctx, executor);

    TFE_DeleteContext(ctx);
    TFE_DeleteExecutor(executor);
  }

  {
    TFE_ContextOptions* opts = TFE_NewContextOptions();

        // parallel device.
        device_(device),
        executor_(
            TFE_NewExecutor(is_async, /*enable_streaming_enqueue=*/true,
                            /*in_flight_nodes_limit=*/in_flight_nodes_limit)),
        op_(nullptr),
        thread_(tensorflow::Env::Default()->StartThread(
            tensorflow::ThreadOptions(), "parallel_device_execute",
            std::bind(&DeviceThread::Run, this))) {}

 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

  // Helper to copy a tensor handle from another device once for each component
  // of the ParallelDevice.
  //

//   materialized in the address space of the calling process.
// * int is used as an index into arrays.
// * Deletion functions are safe to call on nullptr.
//
// Questions left to address:
// * Might at some point need a way for callers to provide their own Env.
// * Maybe add TF_TensorShape that encapsulates dimension info.
//
// Design decisions made:
// * Backing store for tensor memory has an associated deallocation

    TF_Status* status);
// Get the number of dimension of a variable
TF_CAPI_EXPORT extern int TF_CheckpointReaderGetVariableNumDims(
    TF_CheckpointReader* reader, const char* name);
// Load the weight of a variable
TF_CAPI_EXPORT extern TF_Tensor* TF_CheckpointReaderGetTensor(
    TF_CheckpointReader* reader, const char* name, TF_Status* status);

// TF_NewAttrBuilder() returns an object that you can set attributes on as

struct TFE_Executor {
  explicit TFE_Executor(bool async, bool enable_streaming_enqueue,
                        int in_flight_nodes_limit)
      : owned_executor(new tensorflow::EagerExecutor(
            async, enable_streaming_enqueue, in_flight_nodes_limit)) {}

  explicit TFE_Executor(tensorflow::EagerExecutor* executor)
      : owned_executor(nullptr), unowned_executor(executor) {}

// Creates a new eager Executor. Nodes in one executor are guaranteed to be
// executed in sequence. Assigning nodes to different executors allows executing
// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the
// maximum number of in flight async nodes. Enqueuing of additional async ops
// after the limit is reached blocks until some inflight nodes finishes.

//   * `void* ext` is a free-form field that can be populated by
//     a plugin in `TP_*` structs or potential future extension points .
//
// Example usage:
//
//   /* Sample TensorFlow code below, exact implementation might differ. */
//   // Version checking uses `struct_size`. It should be set both by core
//   // and the plugin.
//   TP_OptimizerRegistrationParams params{
//       TP_OPTIMIZER_REGISTRATION_PARAMS_STRUCT_SIZE};

*   `tf.keras`
    *   Export `depthwise_conv2d` in `tf.keras.backend`.
    *   In Keras Layers and Models, Variables in `trainable_weights`,
        `non_trainable_weights`, and `weights` are explicitly deduplicated.
    *   Keras `model.load_weights` now accepts `skip_mismatch` as an argument.
        This was available in external Keras, and has now been copied over to
        `tf.keras`.

  status->status = absl::OkStatus();
}

TFE_Executor* TFE_NewExecutor(bool is_async, bool enable_streaming_enqueue,
                              int in_flight_nodes_limit) {
  return new TFE_Executor(is_async, enable_streaming_enqueue,
                          in_flight_nodes_limit);
}

void TFE_DeleteExecutor(TFE_Executor* executor) { delete executor; }

bool TFE_ExecutorIsAsync(TFE_Executor* executor) {