}
  ~ReluGradientFunction() override {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Unref();
      }
    }
  }

 private:
  // TODO(b/174778737): Only hold needed outputs.
  vector<AbstractTensorHandle*> forward_outputs_;
};

Status BroadcastMul(AbstractContext* ctx, AbstractTensorHandle* vec,
                    AbstractTensorHandle* mat,

  while (*pos < static_cast<size_t>(graph->graph.num_node_ids())) {
    Node* node = graph->graph.FindNodeId(*pos);
    // FindNodeId() returns nullptr for nodes that have been deleted.
    // We aren't currently allowing nodes to be deleted, but it is safer
    // to still check.
    if (node != nullptr) return ToOperation(node);
    *pos += 1;
  }

  // No more nodes.
  return nullptr;
}

  auto* optimizer_options =
      config.mutable_graph_options()->mutable_optimizer_options();
  if (enable) {
    optimizer_options->set_global_jit_level(tensorflow::OptimizerOptions::ON_1);

    // These XLA flags are needed to trigger XLA properly from C (more generally
    // non-Python) clients. If this API is called again with `enable` set to
    // false, it is safe to keep these flag values as is.

  std::vector<AbstractTensorHandle*> temp_outputs(2);
  TF_RETURN_IF_ERROR(
      ops::IdentityN(ctx, inputs, absl::MakeSpan(temp_outputs), "IdentityN"));
  // Although, `ops::IdentityN` returns 2 tensors, the first tensor isn't needed
  // for computing gradient so we could safely drop it.
  outputs[0] = temp_outputs[1];
  temp_outputs[0]->Unref();
  return absl::OkStatus();
}

class CppGradients

namespace {

// Managing context for the DLManagedTensor, will manage the lifetime of
// DLManagedTensor. When calling DLManagedTensor::deleter, it will notify the
// original framework of destruction, and this context will be deleted also.
struct TfDlManagedTensorCtx {
  TensorReference reference;
  std::vector<int64_t> shape;
  std::vector<int64_t> strides;
  DLManagedTensor tensor;

  for (int i = 0; i < *num_retvals; i++) {
    // TODO(srbs): Manage refcount of ForwardOperation's inputs/outputs.
    forward_op_->outputs.push_back(retvals[i]);
  }
  // TODO(b/166669239): This is needed to support AttrBuilder::Get for string
  // attributes. Number type attrs and DataType attrs work fine without this.
  // Consider getting rid of this and making the behavior between number types
  // and string consistent.

      "or NotDifferentiableGradientFunction.",
      s.message());
  ASSERT_EQ(nullptr, outputs[0]);
}

// TODO(b/164171226): Enable this test with tfrt after AddInputList is
// supported. It is needed for IdentityN.
#ifdef PLATFORM_GOOGLE
INSTANTIATE_TEST_SUITE_P(
    UnifiedCAPI, CppGradients,
    ::testing::Combine(::testing::Values("graphdef", "mlir"),
                       /*tfrt*/ ::testing::Values(false),

                      TF_Status* status) {
  GeneralAssignment("AssignVariableOp", context, value, status);
}

// Passed to `TF_NewTensor` to indicate how an array of floats should be
// deleted.
static void FloatDeallocator(void* data, size_t, void* arg) {
  delete[] static_cast<float*>(data);
}

// Creates a TFE_TensorHandle with value `v`.
TensorHandlePtr FloatTensorHandle(float v, TF_Status* status) {

    std::unique_ptr<ParallelTensor> t, TF_Status* status) {
  // The resulting TensorHandle owns an opaque pointer to "device memory", which
  // for a ParallelDevice is really a ParallelTensor. When the TensorHandle is
  // deleted, it will call ParallelTensorDeallocator to free the struct.
  ParallelTensor* t_released = t.release();
  TFE_CustomDeviceTensorHandleMethods handle_methods;
  handle_methods.num_dims = &ParallelTensorNumDims;

  }
  ~MatMulGradientFunction() override {
    for (auto input : forward_inputs_) {
      if (input) {
        input->Unref();
      }
    }
  }

 private:
  // TODO(b/174778737): Only hold needed inputs.
  vector<AbstractTensorHandle*> forward_inputs_;
  AttrBuilder forward_attrs_;
};

class NegGradientFunction : public GradientFunction {
 public:
  Status Compute(AbstractContext* ctx,