#
#   RBE is incompatible with local caching, so you must remove
#   disk_cache, public_cache, and public_cache_push from your $TFCI file.
#
# To use RBE, you must first run `gcloud auth application-default login`, then:
export TFCI=py311,linux_x86,rbe

# Finally: Run your script of choice.
#   If you've clicked on a test result from our CI (via a dashboard or GitHub link),

  # Python 3.12 removed the module `imp` which is needed by gcloud CLI so we set
  # `CLOUDSDK_PYTHON` to Python 3.11 which is the system Python on TFCI Mac
  # VMs.
  export CLOUDSDK_PYTHON=$(which python3.11)
  gcloud auth activate-service-account --key-file="${GOOGLE_APPLICATION_CREDENTIALS}"
fi
set -x

# When cross-compiling with RBE, we need to copy the macOS sysroot to be

# Upload all build artifacts to e.g. gs://tensorflow/versions/2.16.0-rc1 (releases) or
# gs://tensorflow/nightly/2.16.0-dev20240105 (nightly), overwriting previous values.
if [[ "$TFCI_ARTIFACT_FINAL_GCS_ENABLE" == 1 ]]; then
  gcloud auth activate-service-account --key-file="$TFCI_ARTIFACT_FINAL_GCS_SA_PATH"

  # $TF_VER_FULL will resolve to e.g. "2.15.0-rc2". Since $TF_VER_FULL comes
  # from get_versions.sh, which must be run *after* update_version.py, FINAL_URI

  StringVectorToArrays(list, &list_ptrs, &list_lens);
  int list_total_size = 0;
  for (const auto& s : list) {
    list_total_size += s.size();
  }

  auto desc = init("list(string)");
  TF_SetAttrStringList(desc, "v", list_ptrs.get(), list_lens.get(),
                       list.size());

  auto oper = TF_FinishOperation(desc, s_);
  ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);

};

TEST_P(CppGradients, TestReluGrad) {
  status_ = registry_.Register("Relu", ReluRegisterer);
  ASSERT_EQ(errors::OK, status_.code()) << status_.message();

  auto ReluGradModel = BuildGradModel(ReluModel, registry_);

  float X_vals[] = {1.0f, 2.0f, 3.0f, -5.0f, -4.0f, -3.0f, 2.0f, 10.0f, -1.0f};
  int64_t X_dims[] = {3, 3};
  AbstractTensorHandlePtr X;
  {

    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(

  *result = outputs[0];
  return absl::OkStatus();
}
}  // namespace

Status GradientRegistry::Register(
    const string& op_name, GradientFunctionFactory gradient_function_factory) {
  auto iter = registry_.find(op_name);
  if (iter != registry_.end()) {
    const string error_msg = "Gradient already exists for op: " + op_name + ".";
    return errors::AlreadyExists(error_msg);
  }

  if (dlMTensor->deleter != nullptr) {
    dlMTensor->deleter(dlMTensor);
  }
}

void* TFE_HandleToDLPack(TFE_TensorHandle* h, TF_Status* status) {
  auto tf_dlm_context = GetDlContext(h, status);
  if (!status->status.ok()) {
    return nullptr;
  }

  auto* tf_dlm_data = TFE_TensorHandleDevicePointer(h, status);
  if (!status->status.ok()) {
    return nullptr;
  }