using tensorflow::RecordMutation;
  tensorflow::AttrValue attr_val;
  if (!attr_val.ParseFromArray(attr_value_proto->data,
                               attr_value_proto->length)) {
    status->status = absl::InvalidArgumentError("Invalid AttrValue proto");
    return;
  }

  mutex_lock l(graph->mu);
  op->node.AddAttr(attr_name, attr_val);
  RecordMutation(graph, *op, "setting attribute");
}

  bool transpose_a_vals[] = {false, false, true, true};
  bool transpose_b_vals[] = {false, true, false, true};
  float dA_vals[4][9] = {{24, 15, 6, 24, 15, 6, 24, 15, 6},
                         {18, 15, 12, 18, 15, 12, 18, 15, 12},
                         {24, 24, 24, 15, 15, 15, 6, 6, 6},
                         {18, 18, 18, 15, 15, 15, 12, 12, 12}};
  float dB_vals[4][9] = {{12, 12, 12, 15, 15, 15, 18, 18, 18},

        found_dtype = true;
      } else {
        return false;
      }
    } else if (attr.first == "value") {
      if (attr.second.has_tensor() &&
          attr.second.tensor().int_val_size() == 1 &&
          attr.second.tensor().int_val(0) == v) {
        found_value = true;
      } else {
        return false;
      }
    }
  }
  return found_dtype && found_value;
}

  AbstractTensorHandlePtr sum_dims;
  {
    vector<int32_t> vals(num_dims_out);
    int64_t vals_shape[] = {num_dims_out};
    Range(&vals, 0, num_dims_out);
    AbstractTensorHandle* sum_dims_raw = nullptr;
    TF_RETURN_IF_ERROR(TestTensorHandleWithDims<int32_t, TF_INT32>(
        ctx, vals.data(), vals_shape, 1, &sum_dims_raw));
    sum_dims.reset(sum_dims_raw);
  }

    // runs are unlikely to clash.
    std::random_device rd;
    std::mt19937 gen(rd());
    std::uniform_int_distribution<> distribution;
    std::string rng_val = std::to_string(distribution(gen));
    return tensorflow::io::JoinPath(std::string(test_dir), rng_val);
  } else {
    return "";
  }
}

static std::string* GetTmpDir() {
  static std::string tmp_dir = InitializeTmpDir();
  if (tmp_dir == "")

};

TEST_P(GradientCheckerTest, TestMatMul) {
  float A_vals[] = {1.0f, 2.0f, 3.0f, 4.0f};
  int64_t A_dims[] = {2, 2};
  AbstractTensorHandlePtr A;
  {
    AbstractTensorHandle* A_raw;
    Status s = TestTensorHandleWithDims<float, TF_FLOAT>(ctx_.get(), A_vals,
                                                         A_dims, 2, &A_raw);

         TF_TensorByteSize(result_tensor));

  // Build expected result & verify.
  float e_vals[] = {19.0f, 22.0f, 43.0f, 50.0f};

  int data_len = 4;  // length of e_vals
  for (int i = 0; i < data_len; i++) {
    EXPECT_EQ(result_data[i], e_vals[i]);
  }

  TF_DeleteTensor(result_tensor);
  TF_DeleteAbstractTensor(result);
  TF_DeleteOutputList(o);

  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;
  {
    AbstractTensorHandle* X_raw;
    status_ = TestTensorHandleWithDims<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), X_vals, X_dims, 2, &X_raw);
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();

  for (size_t i = 0; i < values.size(); i++) {
    memset(&option_values[i], 0, sizeof(option_values[i]));
    option_values[i].buffer_val.buf = const_cast<char*>(values[i].c_str());
    option_values[i].buffer_val.buf_length = values[i].size();
  }
  option_value.values = &option_values[0];
  option.value = &option_value;
  UniquePtrTo_TF_Status plugin_status(TF_NewStatus(), TF_DeleteStatus);

    //        |         |
    //      Const_0    Const_1
    //
    const float const0_val[] = {1.0, 2.0, 3.0, 4.0};
    const float const1_val[] = {1.0, 0.0, 0.0, 1.0};
    TF_Operation* const0 = FloatConst2x2(graph_, s_, const0_val, "Const_0");
    TF_Operation* const1 = FloatConst2x2(graph_, s_, const1_val, "Const_1");
    TF_Operation* matmul = MatMul(graph_, s_, const0, const1, "MatMul");