Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // Grad for Softmax Input
    TF_RETURN_IF_ERROR(BroadcastMul(
        ctx, grad_outputs[0], forward_outputs_[1],
        grad_inputs.subspan(0, 1)));  // upstream_grad * local softmax grad

    // Grad for labels is null

    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?
      if (grad_input) {
        grad_input->Ref();
      }
      grad_inputs[i] = grad_input;
    }
    return absl::OkStatus();
  }
  ~IdentityNGradientFunction() override {}
};

    if (grad_inputs[0]) {
      grad_inputs[0]->Ref();
    }
    return absl::OkStatus();
  }
};

// Computes:
//
// @tf.custom_gradient
// def f(input):
//   def grad(grads):
//     return grads[0]
//   return tf.exp(input), grad
// outputs = [f(inputs[0])]
Status ExpWithPassThroughGrad(AbstractContext* ctx,
                              absl::Span<AbstractTensorHandle* const> inputs,

                                       bool has_packed_input = false) {
  return TestRemoteExecuteSilentCopies(async, remote, /*func=*/true,
                                       heavy_load_on_streaming_rpc,
                                       remote_func_outputs, has_packed_input);
}

TEST(CAPI, RemoteExecuteSilentCopiesAsyncFunc) {

    if (!outputs[i]) continue;

    AbstractTensorHandlePtr numerical_grad;
    {
      AbstractTensorHandle* numerical_grad_raw;
      s = CalcNumericalGrad(ctx, model, inputs,
                            /*input_index=*/i, use_function,
                            &numerical_grad_raw);
      ASSERT_EQ(errors::OK, s.code()) << s.message();
      numerical_grad.reset(numerical_grad_raw);
    }

    TF_Tensor* numerical_tensor;

    TFE_TensorHandle* input = TFE_OpGetFlatInput(original_op, j, s);
    if (TF_GetCode(s) != TF_OK) return;
    const char* input_device = TFE_TensorHandleDeviceName(input, s);
    if (TF_GetCode(s) != TF_OK) return;
    if (dev->device_name == input_device) {
      LoggedTensor* t = reinterpret_cast<LoggedTensor*>(
          TFE_TensorHandleDevicePointer(input, s));
      if (TF_GetCode(s) != TF_OK) return;

                         int input_index, bool use_function,
                         AbstractTensorHandle** numerical_grad) {
  vector<AbstractTensorHandle*> theta_inputs(inputs.size());
  for (int i{}; i < inputs.size(); ++i) {
    theta_inputs[i] = inputs[i];
  }

  AbstractTensorHandle* theta =
      theta_inputs[input_index];  // parameter we are grad checking

      absl::StrCat("    signature {"
                   "      name: 'MatMulFunction'"
                   "      input_arg {"
                   "        name: 'a'"
                   "        type: DT_FLOAT"
                   "      }"
                   "      input_arg {"
                   "        name: 'b'"
                   "        type: DT_FLOAT"
                   "      }"

                             std::vector<AbstractTensorHandle*>* params) {
  tracing::TracingTensorHandle* handle = nullptr;
  for (auto input : inputs) {
    PartialTensorShape shape;
    TF_RETURN_IF_ERROR(input->Shape(&shape));
    TF_RETURN_IF_ERROR(dyn_cast<tracing::TracingContext>(ctx)->AddParameter(
        input->DataType(), shape, &handle));
    params->emplace_back(handle);
  }
  return absl::OkStatus();
}

    Model model, AbstractContext* ctx,
    absl::Span<AbstractTensorHandle* const> inputs, int input_index,
    float* expected_grad, int num_grad, bool use_function,
    double abs_error = 1e-2) {
  Status s;
  AbstractTensorHandlePtr numerical_grad;
  {
    AbstractTensorHandle* numerical_grad_raw;
    s = CalcNumericalGrad(ctx, model, inputs, input_index, use_function,
                          &numerical_grad_raw);