// devices.
  virtual Status Dim(int dim_index, int64_t* dim) const = 0;

  // Returns the device which created the handle.
  virtual const char* DeviceName(Status* status) const = 0;
  // Returns the device where the tensor was placed.
  virtual const char* BackingDeviceName(Status* status) const = 0;
  // Returns the device type which created the handle.

  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());

// Controls how to act when we try to run an operation on a given device but
// some input tensors are not on that device.
// LINT.IfChange
// Note: Keep in sync with internal copy of enum in eager/context.h.
typedef enum TFE_ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  TFE_DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  TFE_DEVICE_PLACEMENT_WARN = 1,

  // returned by DeviceName should be "/device:GPU:*" until a particular GPU is
  // chosen for the operation by the device placement logic in the
  // executor. After that, the value returned by DeviceName will be a full
  // device name such as "/job:localhost/replica:0/task:0/device:GPU:1".
  virtual const string& DeviceName() const = 0;

  // Sets the operation device name.
  //

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  const DLTensor* dltensor_out = &dlm_out->dl_tensor;
  EXPECT_EQ(dltensor_out->device.device_type, dltensor_in->device.device_type);
  EXPECT_EQ(dltensor_out->device.device_id, dltensor_in->device.device_id);
  EXPECT_EQ(dltensor_out->ndim, dltensor_in->ndim);
  EXPECT_EQ(dltensor_out->dtype.code, dltensor_in->dtype.code);

    if _os.path.exists(_plugin_dir):
      _ll.load_library(_plugin_dir)
      # Load Pluggable Device Library
      _ll.load_pluggable_device_library(_plugin_dir)

if _os.getenv("TF_PLUGGABLE_DEVICE_LIBRARY_PATH", ""):
  _ll.load_pluggable_device_library(
      _os.getenv("TF_PLUGGABLE_DEVICE_LIBRARY_PATH")
  )

# Add Keras module aliases

}

void TF_SetRequestedDevice(TF_Graph* graph, TF_Operation* op,
                           const char* device) {
  using tensorflow::RecordMutation;
  mutex_lock l(graph->mu);
  op->node.set_requested_device(device);
  RecordMutation(graph, *op, "setting device");
}

void TF_RemoveAllControlInputs(TF_Graph* graph, TF_Operation* op) {
  mutex_lock l(graph->mu);

void TF_AbstractOpSetOpType(TF_AbstractOp* op, const char* const op_type,
                            TF_Status* s) {
  tsl::Set_TF_Status_from_Status(
      s, unwrap(op)->Reset(op_type,
                           /*raw_device_name=*/nullptr));
}

void TF_AbstractOpSetOpName(TF_AbstractOp* op, const char* const op_name,
                            TF_Status* s) {
  TracingOperation* tracing_op = dyn_cast<TracingOperation>(unwrap(op));

  std::unique_ptr<ParallelTensor> DeviceIDs(TFE_Context* context,
                                            TF_Status* status) const;

  // The number of devices operations run on.
  size_t num_underlying_devices() const { return underlying_devices_.size(); }

  // The devices operations run on.
  const std::vector<std::string>& underlying_devices() const {
    return underlying_devices_;

ParallelDevice::ParallelDevice(const std::vector<std::string>& devices,
                               bool is_async, int in_flight_nodes_limit)
    : underlying_devices_(devices),
      default_cancellation_manager_(absl::make_unique<CancellationManager>()) {
  device_threads_.reserve(devices.size());
  for (int device_index = 0; device_index < devices.size(); ++device_index) {
    device_threads_.emplace_back(new DeviceThread(