ASSERT_EQ(errors::OK, status_.code()) << status_.message();
    ASSERT_NO_FATAL_FAILURE(CheckTensorValue(outputs[0], dA_vals[i],
                                             /*dims*/ {3, 3},
                                             /*abs_error*/ 0));
    ASSERT_NO_FATAL_FAILURE(CheckTensorValue(outputs[1], dB_vals[i],
                                             /*dims*/ {3, 3},

      context.get(), in_components, device_name, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Loop to make synchronization failures more deterministic
  for (int i = 0; i < 100; ++i) {
    TensorHandlePtr multiply_result(
        Multiply(context.get(), combined_value.get(), combined_value.get(),
                 status.get()));

               {x1.get(), x2.get()}, absl::MakeSpan(outputs), UseFunction());
  ASSERT_EQ(errors::OK, status_.code()) << status_.message();
  EXPECT_EQ(outputs[0], nullptr);
  ASSERT_NO_FATAL_FAILURE(CheckTensorValue(outputs[1], {1.0f}, /*dims*/ {},
                                           /*abs_error*/ 0));
  outputs[1]->Unref();
}

#ifdef PLATFORM_GOOGLE
INSTANTIATE_TEST_SUITE_P(

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // TODO(b/136478427): Figure out how to correctly shut the server down.
  worker_server.release();

  // Update the server def with a new set of names (worker instead of
  // localhost).

                         tensorflow::FunctionLibraryDefinition* flib_def,
                         std::vector<std::string>* control_ret_node_names,
                         bool* control_rets_updated) override {
    // Inject failure to function instantiation if finding a node that contains
    // the given node name (error_node_) and requested device (error_device_).
    for (const auto node : graph->get()->nodes()) {

    // would otherwise deadlock).
    if (TF_GetCode(status) != TF_OK &&
        (first_bad_status == nullptr
         // Prefer propagating non-cancellation related statuses to avoid
         // shadowing the original failure.
         || TF_GetCode(first_bad_status.get()) == TF_CANCELLED)) {
      first_bad_status.reset(TF_NewStatus());
      TF_SetStatus(first_bad_status.get(), TF_GetCode(status),
                   TF_Message(status));

    ASSERT_EQ(errors::OK, s.code()) << s.message();
    B.reset(B_raw);
  }

  float expected_dA[4] = {-.5f, 2.0f, -.5f, 2.0f};
  ASSERT_NO_FATAL_FAILURE(CompareNumericalAndManualGradients(
      MatMulModel, ctx_.get(), {A.get(), B.get()}, 0, expected_dA, 4,
      UseFunction()));
}

TEST_P(GradientCheckerTest, TestMul) {
  AbstractTensorHandlePtr x;
  {

        immediate_execution_ctx_.get(), X_vals, X_dims, 2, &X_raw);
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();
    X.reset(X_raw);
  }

  ASSERT_NO_FATAL_FAILURE(CompareNumericalAndAutodiffGradients(
      ReluModel, ReluGradModel, immediate_execution_ctx_.get(), {X.get()},
      UseFunction()));

  // Mathematically, Relu isn't differentiable at `0`. So `gradient_checker`

              const std::vector<TF_Operation*>& outputs,
              const std::vector<string>& output_names,
              bool expect_failure = false) {
    DefineT(num_opers, opers, ToOutput(inputs), ToOutput(outputs), output_names,
            expect_failure);
  }

  // Caller must delete[] the returned value
  static const char** ToArray(const std::vector<string>& strs) {
    const char** ptr = nullptr;

      ADD_FAILURE() << i << " -- " << TF_Message(status.get());
      continue;
    }
    auto tag = tensorflow::strings::StrCat("Device #", i, " (", name, ")");
    // Copy to device
    TFE_TensorHandle* hdevice =
        TFE_TensorHandleCopyToDevice(hcpu, ctx, name.c_str(), status.get());
    if (TF_GetCode(status.get()) != TF_OK) {
      ADD_FAILURE() << tag << " -- " << TF_Message(status.get());