tensorflow::CustomDevice* device,
                               tensorflow::DataType dtype, void* data,
                               TFE_CustomDeviceTensorHandleMethods methods)
      : tensorflow::CustomDeviceTensorHandle(context, device, dtype),
        data_(data),
        methods_(methods) {}

  ~CAPICustomDeviceTensorHandle() override { methods_.deallocator(data_); }

      "      attr {"
      "        key: 'dtype'"
      "        value {"
      "          type: DT_FLOAT"
      "        }"
      "      }"
      "    }"
      "    node_def {"
      "      name: 'read1'"
      "      op: 'ReadVariableOp'"
      "      input: 'var'"
      "      device: '/job:localhost/replica:0/task:1/device:CPU:0'"
      "      attr {"
      "        key: 'dtype'"
      "        value {"

  /// @return The unwrapped value. If this `TaggedValue` type does not currently
  ///         contain a value of type `T`, the program terminates via a call to
  ///         `assert`.
  template <typename T>
  T& get() {
    assert(type_ == EnumValueOf<T>::value);
    return UnionAccess<T>::unsafe_reference(*this);
  }

  /// @brief Get the underlying value based on type.
  ///
  /// @tparam T The desired return type.

  FunctionDef make_ref_float = FunctionDefHelper::Define(
      "RefFloatFn", {}, {"r:float"}, {},
      {{{"var"},
        "VariableV2",
        {},
        {{"dtype", DT_FLOAT}, {"shape", TensorShape({})}}},
       {{"r"}, "Identity", {"var"}, {{"T", DT_FLOAT}}}});
  *fdef_lib->add_function() = make_ref_float;
}

void AddRegularFunctionFunctionDef(FunctionDefLibrary* fdef_lib) {

    size_t idx = type_and_idx.index();
    auto result_ty = mlir::cast<mlir::RankedTensorType>(type_and_idx.value());

    // If the result type isn't static, then the owner of the result may be a
    // cast op from a more specific bounded type to an unbounded dynamic type.
    // Use the bounded type to get the buffer size.
    mlir::RankedTensorType buffer_ty = result_ty;
    if (!buffer_ty.hasStaticShape()) {

  TF_Status* status = TF_NewStatus();
  // Create the variable handle.
  TFE_Op* op = TFE_NewOp(ctx, "VarHandleOp", status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpSetAttrShape(op, "shape", {}, 0, status);
  TFE_OpSetAttrString(op, "container", "localhost", 0);
  TFE_OpSetAttrString(op, "shared_name", "", 0);
  if (!device_name.empty()) {

      tensorflow::OpInfo::TensorProperties in_props;
      Status s = tensorflow::BufferToMessage(in_props_buf[0], &in_props);
      TF_ASSERT_OK(s);

      EXPECT_EQ(DT_FLOAT, in_props.dtype());
      EXPECT_FALSE(in_props.shape().unknown_rank());
      EXPECT_EQ(2, in_props.shape().dim_size());
      EXPECT_EQ(10, in_props.shape().dim(0).size());
      EXPECT_EQ(1, in_props.shape().dim(1).size());

  }
  strings::StrAppend(&ret, "}");
  return ret;
}

string PrintTensor(const TensorProto& tensor_proto) {
  Tensor t(tensor_proto.dtype());
  CHECK(t.FromProto(tensor_proto));
  const int64_t num_elts = t.NumElements();
  switch (t.dtype()) {
    case DT_FLOAT:
      return PrintArray(num_elts, t.flat<float>().data());
    case DT_DOUBLE:
      return PrintArray(num_elts, t.flat<double>().data());

  const tensorflow::Tensor& arg_tensor = cc_ctx->input(index);
  absl::Status cc_status;
  if (arg_tensor.dtype() != tensorflow::DT_RESOURCE) {
    cc_status = absl::InvalidArgumentError(
        absl::StrCat("Trying to obtain resource handle from Input[", index,
                     "], which is not type DT_RESOURCE."));
    status->status = cc_status;
    return nullptr;
  }
  const tensorflow::ResourceHandle& handle =

  }

  // Derive the output types. The result type is derived by using the
  // attributes attched to the result type of the signature. The attribute
  // value should be either in the attribute argument list or the derived
  // attribute from the input tensors. All the result type
  // are unranked, and shape inference should be applied afterwards.
  SmallVector<Type, 4> output_types;