CHECK_LT(output.index, ic->num_outputs());  // Crash OK
    const auto* shapes_and_types =
        ic->output_handle_shapes_and_types(output.index);
    if (shapes_and_types == nullptr) return nullptr;

    for (const auto& p : *shapes_and_types) {
      auto* out_shape_and_type = handle_data.add_shape_and_type();
      ic->ShapeHandleToProto(p.shape, out_shape_and_type->mutable_shape());

namespace gradients {
namespace {

class ReluGradientFunction : public GradientFunction {
 public:
  explicit ReluGradientFunction(vector<AbstractTensorHandle*> f_outputs)
      : forward_outputs_(f_outputs) {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,

         (GetFactories()[name] == factory) &&
             "Duplicate tracing factory registration");
  GetFactories()[name] = factory;
}

Status SetDefaultTracingEngine(const char* name) {
  auto entry = GetFactories().find(name);
  if (entry != GetFactories().end()) {
    default_factory = GetFactories().find(name)->second;
    return absl::OkStatus();
  }
  string msg = absl::StrCat(

  ASSERT_NE(handle, nullptr)
      << TF_Message(status) << " (shape=[" << absl::StrJoin(shape, ",")
      << "], strides=[" << absl::StrJoin(strides, ",") << "])";

  auto* dlm_out =
      static_cast<DLManagedTensor*>(TFE_HandleToDLPack(handle, status));
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  const DLTensor* dltensor_out = &dlm_out->dl_tensor;

  TF_Tensor* analytical_tensor;
  auto s = GetValue(t, &analytical_tensor);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

  int64_t num_elem_analytical = 1;
  auto num_dims_analytical = TF_NumDims(analytical_tensor);
  ASSERT_EQ(dims.size(), num_dims_analytical);
  for (int j = 0; j < num_dims_analytical; j++) {
    auto dim_analytical = TF_Dim(analytical_tensor, j);

  tensorflow::mutex_lock c(graph->mu);
  const auto& debug_str = graph->graph.ToGraphDefDebug().DebugString();
  *len = debug_str.size();
  char* ret = static_cast<char*>(malloc(*len + 1));
  memcpy(ret, debug_str.c_str(), *len + 1);
  return ret;
}

char* TF_FunctionDebugString(TF_Function* func, size_t* len) {
  const auto& debug_str = DebugString(func->record->fdef());
  *len = debug_str.size();

} PosixFile;

static void Cleanup(TF_WritableFile* file) {
  auto posix_file = static_cast<PosixFile*>(file->plugin_file);
  plugin_memory_free(const_cast<char*>(posix_file->filename));
  delete posix_file;
}

static void Append(const TF_WritableFile* file, const char* buffer, size_t n,
                   TF_Status* status) {
  auto posix_file = static_cast<PosixFile*>(file->plugin_file);

  TFE_OpSetAttrShape(handle_op.get(), "shape", /*dims=*/nullptr, /*num_dims=*/0,
                     status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  auto outputs =
      parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),
                              "VarHandleOp", TFE_OpGetAttrs(handle_op.get()),

                 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) {
        grad_input->Ref();
      }

  ASSERT_EQ(errors::OK, s.code()) << s.message();

  TF_Tensor* result_tensor;
  s = GetValue(outputs[0], &result_tensor);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto result_value = static_cast<float*>(TF_TensorData(result_tensor));
  EXPECT_EQ(*result_value, 1.0);
  outputs[0]->Unref();
  TF_DeleteTensor(result_tensor);
  result_tensor = nullptr;
}

INSTANTIATE_TEST_SUITE_P(