tensorflow::condition_variable finished_join_;

  // Temporary state between `StartExecute` and `Join`.
  //
  //   Inputs; pointers are to objects not owned by the DeviceThread, but which
  //   are expected to live at least until `Join` finishes:
  TFE_Context* context_ TF_GUARDED_BY(execution_mutex_);
  const char* operation_name_ TF_GUARDED_BY(execution_mutex_);

// registration.
//
// Currently this is an error, and un-packing ParallelTensors must be performed
// explicitly by running a TPUReplicatedOutput operation on the parallel device.
//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either
// need to return something here or address these use-cases one by one.

  for (int i{}; i < inputs.size(); ++i) {
    theta_inputs[i] = inputs[i];
  }

  AbstractTensorHandle* theta =
      theta_inputs[input_index];  // parameter we are grad checking

  // Convert from AbstractTensor to TF_Tensor.
  TF_Tensor* theta_tensor;
  TF_RETURN_IF_ERROR(GetValue(theta, &theta_tensor));

  // Get number of elements and fill data.

  // Set to true whenever a TensorHandle is copied onto the device
  bool* arrived_flag;
  // Set to true whenever an operation is executed
  bool* executed_flag;
  // If true, only explicit op placements are accepted. If false, uses
  // type-based dispatch.
  bool strict_scope_placement;
};

struct LoggedTensor {
  TFE_TensorHandle* tensor;
  LoggedTensor() = delete;

  // Pack 3 variable handles into one TFE_TensorHandle.
  // When remote is false, function device is placed on task0. Handle types are
  // REMOTE, REMOTE, LOCAL on task0. When remote is true, function device is
  // placed on task1, Handle types are LOCAL, REMOTE, LOCAL on task1.
  int num_replicas = 3;
  std::vector<TFE_TensorHandle*> handles = {h0, h1, h2};
  TFE_TensorHandle* packed_handle =

    if (TF_GetCode(status) != TF_OK) {
      TF_DeleteStatus(status);
      return "";
    }
    TF_DeleteStatus(status);

    // We add a random value into `test_dir` to ensures that two consecutive
    // runs are unlikely to clash.
    std::random_device rd;
    std::mt19937 gen(rd());
    std::uniform_int_distribution<> distribution;
    std::string rng_val = std::to_string(distribution(gen));

#include "tensorflow/c/tf_status_internal.h"
#include "tensorflow/core/lib/core/status_test_util.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {

#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {

                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    /* Given upstream grad U and a matmul op A*B, the gradients are:
     *
     *    dA = U * B.T
     *    dB = A.T * U
     *
     *    where A.T means `transpose(A)`
     */
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

    // Get transpose attrs

//
// This has three benefits:
//   * allows having newer plugins than the current core TensorFlow: the
//     additional entries in the plugin's table are just discarded;
//   * allows having older plugins than the current core TensorFlow (though
//     we are still warning users): the entries that core TensorFlow expects
//     but plugins didn't provide will be set to `nullptr` values and core