}
  tsl::Set_TF_Status_from_Status(s, tracing_op->SetOpName(op_name));
}

void TF_AbstractOpSetAttrType(TF_AbstractOp* op, const char* const attr_name,
                              TF_DataType value, TF_Status* s) {
  Status status =
      unwrap(op)->SetAttrType(attr_name, static_cast<DataType>(value));
  TF_SetStatus(s, static_cast<TF_Code>(status.code()),
               tsl::NullTerminatedMessage(status));
}

};

// TODO(vnvo2409): Add python test
class BiasAddGradientFunction : public GradientFunction {
 public:
  explicit BiasAddGradientFunction(AttrBuilder f_attrs)
      : forward_attrs_(f_attrs) {}

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {

  forward_op_.attrs.Set(attr_name, value);
  return parent_op_->SetAttrFloat(attr_name, value);
}
Status TapeOperation::SetAttrBool(const char* attr_name, bool value) {
  forward_op_.attrs.Set(attr_name, value);
  return parent_op_->SetAttrBool(attr_name, value);
}
Status TapeOperation::SetAttrType(const char* attr_name, DataType value) {
  forward_op_.attrs.Set(attr_name, value);

                   "      op: 'MatMul'"
                   "      input: 'a'"
                   "      input: 'b'"
                   "      device: '",
                   matmul_device, "'",
                   "      attr {"
                   "        key: 'T'"
                   "        value {"
                   "          type: DT_FLOAT"
                   "        }"
                   "      }"
                   "    }"

              &num_retvals, &forward_op, tape.get(), registry);
  ASSERT_EQ(errors::OK, s.code()) << s.message();

  string read_message;
  s = forward_op.attrs.Get("message", &read_message);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  ASSERT_EQ(read_message, message);
}

Status RecordOperationWithNullGradientFunctionModel(