/*expected_shape*/ make_shape({1, 6}));
  TFE_DeleteOp(reshape_op);
  reshape_op = nullptr;

  TFE_Op* fill_op = TFE_NewOp(tfe_context_, "Fill", status_);
  CHECK_EQ(TF_OK, TF_GetCode(status_)) << TF_Message(status_);
  TFE_OpSetAttrType(fill_op, "T", TF_FLOAT);
  TFE_OpSetAttrType(fill_op, "Tshape", TF_INT32);

  float five = 5.0;

}  // namespace

GradientFunction* AddRegisterer(const ForwardOperation& op) {
  return new AddGradientFunction;
}

GradientFunction* ExpRegisterer(const ForwardOperation& op) {
  return new ExpGradientFunction(op.outputs[0]);
}

GradientFunction* MatMulRegisterer(const ForwardOperation& op) {
  return new MatMulGradientFunction(op.inputs, op.attrs);
}

GradientFunction* SqrtRegisterer(const ForwardOperation& op) {

    auto op_id_it = tensor_tape.find(tensor_id);
    if (op_id_it == tensor_tape.end()) {
      continue;
    }
    int64_t op_id = op_id_it->second;
    auto op_it = op_tape->find(op_id);
    auto result_op_it = result.op_tape.find(op_id);
    if (op_id == -1 || op_it == op_tape->end() ||
        result_op_it != result.op_tape.end()) {
      continue;
    }
    CHECK(result.op_tape.emplace(op_id, op_it->second).second);

        "//tensorflow/core:core_cpu_internal",
        "//tensorflow/core:direct_session",
        "//tensorflow/core:framework",
        "//tensorflow/core:framework_internal",
        "//tensorflow/core:functional_ops_op_lib",
        "//tensorflow/core:lib",
        "//tensorflow/core:math_ops_op_lib",
        "//tensorflow/core:nn_ops_op_lib",
        "//tensorflow/core:no_op_op_lib",

  const std::vector<std::string>& op_names =
      tensorflow::unwrap(ctx)->GetLoggedOpsTestonly();

  std::ostringstream op_names_oss;
  for (const auto& op : op_names) {
    op_names_oss << op << ", ";
  }
  const std::string& op_names_str = op_names_oss.str();
  void* data = tensorflow::port::Malloc(op_names_str.length());
  op_names_str.copy(static_cast<char*>(data), op_names_str.length(), 0);
  buf->data = data;

  NodeDef node_def;
  tensorflow::ImmediateExecutionOperation* op = tensorflow::unwrap(tfe_op);
  node_def.set_name(op->Name());
  node_def.set_op(op->Name());
  for (int i = 0; i < num_inputs; ++i) {
    node_def.add_input("dummy_input");
  }
  OperationFromInterface(op)->Attrs().FillAttrValueMap(node_def.mutable_attr());

  const tensorflow::OpRegistrationData* op_reg_data;
  status->status =

  ParallelDevice parallel_device(std::move(devices));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> handle_op(
      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_OpSetAttrType(handle_op.get(), "dtype", TF_FLOAT);
  TFE_OpSetAttrShape(handle_op.get(), "shape", /*dims=*/nullptr, /*num_dims=*/0,
                     status.get());

  TF_Status* status = TF_NewStatus();
  TFE_Op* op = TFE_NewOp(ctx, "ReadVariableOp", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpAddInput(op, var, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  int num_retvals = 1;
  TFE_Execute(op, out_value, &num_retvals, status);

// TF_NewAttrBuilder() returns an object that you can set attributes on as
// though it were an op. This allows querying properties of that op for
// type-checking purposes like if the op will run on a particular device type.
typedef struct TF_AttrBuilder TF_AttrBuilder;
TF_CAPI_EXPORT extern TF_AttrBuilder* TF_NewAttrBuilder(const char* op_name);
TF_CAPI_EXPORT extern void TF_DeleteAttrBuilder(TF_AttrBuilder* builder);

you want to debug a kernel, you can compile specific files with `-g` using the
`--per_file_copt` bazel option. For example, if you want to debug the Identity
op, which are in files starting with `identity_op`, you can run

```bash
bazel build --config=dbg --per_file_copt=+tensorflow/core/kernels/identity_op.*@-g //tensorflow/tools/pip_package:build_pip_package
```