metadata.list_size = attr->list().field##_size(); \
    __VA_ARGS__;                                      \
    break;                                            \
  }

      LIST_CASE(
          s, TF_ATTR_STRING, metadata.total_size = 0;
          for (int i = 0; i < attr->list().s_size();
               ++i) { metadata.total_size += attr->list().s(i).size(); });
      LIST_CASE(i, TF_ATTR_INT);

  unwrap(o)->expected_num_outputs = num_outputs;
  unwrap(o)->outputs.clear();
  unwrap(o)->outputs.resize(num_outputs);
}
int TF_OutputListNumOutputs(TF_OutputList* o) {
  return unwrap(o)->outputs.size();
}
TF_AbstractTensor* TF_OutputListGet(TF_OutputList* o, int i) {
  return wrap(unwrap(o)->outputs[i]);
}
void TF_OutputListPushBack(TF_OutputList* o, TF_AbstractTensor* tensor,

#include "tensorflow/c/tf_status.h"

// Implementation of a filesystem for POSIX environments.
// This filesystem will support `file://` and empty (local) URI schemes.

static void* plugin_memory_allocate(size_t size) { return calloc(1, size); }
static void plugin_memory_free(void* ptr) { free(ptr); }

// SECTION 1. Implementation for `TF_RandomAccessFile`
// ----------------------------------------------------------------------------

      static_cast<bool>(should_enable);
}

void TF_SetXlaMinClusterSize(int size) {
  tensorflow::MarkForCompilationPassFlags* flags =
      tensorflow::GetMarkForCompilationPassFlags();
  flags->tf_xla_min_cluster_size = size;
}

TF_Buffer* TF_CreateConfig(unsigned char enable_xla_compilation,
                           unsigned char gpu_memory_allow_growth,

             AbstractContext* ctx,
             absl::Span<AbstractTensorHandle* const> inputs,
             absl::Span<AbstractTensorHandle*> outputs) -> Status {
    Tape tape(/*persistent=*/false);
    for (size_t i{}; i < inputs.size(); ++i) {
      tape.Watch(inputs[i]);
    }
    std::vector<AbstractTensorHandle*> temp_outputs(1);
    AbstractContextPtr tape_ctx(new TapeContext(ctx, &tape, grad_registry));
    TF_RETURN_IF_ERROR(

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    CHECK_EQ(grad_outputs.size(), 1);
    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

 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?
      if (grad_input) {

  std::vector<int64_t> input_ids(inputs.size());
  std::vector<tensorflow::DataType> input_dtypes(inputs.size());
  for (int i = 0; i < inputs.size(); i++) {
    input_ids[i] = ToId(inputs[i]);
    input_dtypes[i] = inputs[i]->DataType();
  }
  std::vector<TapeTensor> tape_tensors;
  tape_tensors.reserve(outputs.size());
  for (auto t : outputs) {
    tape_tensors.push_back(TapeTensor(t));
  }

  }
}

// Wraps the deleter function of DLManagedTensor to match the function signature
// TFE_NewTensorHandleFromDeviceMemory.
void DeallocatorWrapperFunc(void* data, size_t len, void* dlmt_vptr) {
  TFE_CallDLManagedTensorDeleter(dlmt_vptr);
}

// Checks whether the stride array matches the layout of compact, row-majored
// data.

  for (int device_index = 0; device_index < underlying_devices_.size();
       ++device_index) {
    DeviceThread* device_thread = device_threads_[device_index].get();
    std::vector<TFE_TensorHandle*> device_inputs;
    device_inputs.reserve(inputs.size());
    for (int input_index = 0; input_index < inputs.size(); ++input_index) {
      // Parallel tensors are divided between operations by device.