CHECK_EQ(grad_inputs.size(), 1);
    grad_inputs[0] = grad_outputs[0];
    if (grad_inputs[0]) {
      grad_inputs[0]->Ref();
    }
    return absl::OkStatus();
  }
};

// Computes:
//
// @tf.custom_gradient
// def f(input):
//   def grad(grads):
//     return grads[0]
//   return tf.exp(input), grad
// outputs = [f(inputs[0])]
Status ExpWithPassThroughGrad(AbstractContext* ctx,

                           void** device_info) {
  TFE_CustomDevice* custom_device = new TFE_CustomDevice;
  custom_device->copy_tensor_to_device = &CopyToLoggingDevice;
  custom_device->copy_tensor_from_device = &CopyTensorFromLoggingDevice;
  custom_device->delete_device = &DeleteLoggingDevice;
  custom_device->execute = &LoggingDeviceExecute;
  *device = custom_device;
  LoggingDevice* logging_device = new LoggingDevice;

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  const char* custom0 = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  const char* custom1 = "/job:localhost/replica:0/task:0/device:CUSTOM:1";
  bool arrived = false;
  bool executed = false;
  RegisterLoggingDevice(context.get(), custom0,
                        /*strict_scope_placement=*/false, &arrived, &executed,

  const char* first_device = "/job:worker/replica:0/task:1/device:CPU:0";
  const char* second_device = "/job:worker/replica:0/task:2/device:CPU:0";
  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  std::array<const char*, 2> underlying_devices{first_device, second_device};
  RegisterParallelDevice(context.get(), device_name, underlying_devices,
                         status.get());

  // additional CPU.
  const char* second_device_name =
      "/job:localhost/replica:0/task:0/device:CUSTOM:1";
  std::array<const char*, 2> second_underlying_devices{
      "/job:localhost/replica:0/task:0/device:CUSTOM:0",
      "/job:localhost/replica:0/task:0/device:CPU:2"};
  RegisterParallelDevice(context.get(), second_device_name,

                             const char* second_device) {
  // Register the custom device
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  const char* device_name = "/job:localhost/replica:0/task:0/device:CUSTOM:0";
  std::array<const char*, 2> underlying_devices{first_device, second_device};

  for (int component_index = 0; component_index < tensors_.size();
       ++component_index) {
    // TODO(allenl): Add a C API for summarizing tensors. Currently custom
    // devices limiting themselves to a C API (for ABI compatibility) would need
    // to implement summarization for component tensors themselves.
    ImmediateExecutionTensorHandle* component =

#include "tensorflow/core/grappler/costs/graph_properties.h"
#include "tensorflow/core/grappler/costs/op_performance_data.pb.h"
#include "tensorflow/core/grappler/grappler_item.h"
#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/status.h"
#include "tensorflow/core/protobuf/rewriter_config.pb.h"
#include "tsl/platform/env.h"

  TF_DeleteStatus(status);
  TFE_DeleteOp(var_op);
  TFE_DeleteOp(var_op_2);
  TFE_DeleteContext(ctx);
}

// Needs to work with a const TFE_Op since custom devices should not modify the
// op they are called with.
TFE_Op* CloneOp(const TFE_Op* other) {
  TF_Status* status = TF_NewStatus();
  TFE_Context* context = TFE_OpGetContext(other, status);

#include "tensorflow/core/common_runtime/eager/attr_builder.h"
#include "tensorflow/core/common_runtime/eager/context.h"
#include "tensorflow/core/common_runtime/eager/custom_device.h"
#include "tensorflow/core/common_runtime/eager/custom_device_op_handler.h"
#include "tensorflow/core/common_runtime/eager/execute.h"
#include "tensorflow/core/common_runtime/eager/placement_utils.h"