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,

        ":not_differentiable",
        "//tensorflow/c/eager:abstract_context",
        "//tensorflow/c/eager:gradients_internal",
    ],
)

tf_cuda_cc_test(
    name = "custom_gradient_test",
    size = "small",
    srcs = [
        "custom_gradient_test.cc",
    ],
    args = ["--heap_check="],  # TODO(b/174752220): Remove
    tags = tf_cuda_tests_tags(),
    deps = [
        "//tensorflow/c:tf_status_helper",

computation, shared state, and the operations that mutate that state. It maps the nodes of a dataflow graph across many machines in a cluster, and within a machine across multiple computational devices, including multicore CPUs, general purpose GPUs, and custom-designed ASICs known as Tensor Processing Units (TPUs). This architecture gives flexibility to the application developer, whereas in previous “parameter server” designs the management of shared state is built into the system, TensorFlow enables developers...

                           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;

// on each underlying device.
//
// For example if `device_name` is
//   "/job:localhost/replica:0/task:0/device:CUSTOM:0"
// and `underlying_devices` is
//   {"/job:localhost/replica:0/task:0/device:GPU:0",
//    "/job:localhost/replica:0/task:0/device:GPU:1"}
// Then executing an operation on CUSTOM:0 will execute it on GPU:0 and GPU:1.
//
// Implicit copies onto `device_name` are allowed, replicating the value once

  // 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,

    id: tfversion
    attributes:
      label: TensorFlow version
      placeholder: e.g., tf 2.8
    validations:
      required: true
  - type: dropdown
    id: Code
    attributes:
      label: Custom code
      options:
        - "Yes"
        - "No"
    validations:
      required: true
  - type: input
    id: OS
    attributes:
      label: OS platform and distribution

                             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};

// Creates a new operation - see `TF_NewOperation` for more details.
//
// The lock for `graph` must be held when calling this function.
//
// Unless implementing advanced behavior, like custom gradient functions, you
// most likely need to call `TF_NewOperation` instead.
TF_CAPI_EXPORT extern TF_OperationDescription* TF_NewOperationLocked(
    TF_Graph* graph, const char* op_type, const char* oper_name);

  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 =