output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    AbstractTensorHandle* upstream_grad = grad_outputs[0];
    AbstractTensorHandle* activations = forward_outputs_[0];

    // Calculate Grad
    std::string name = "relu_grad";

 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    // TODO(b/161805092): Support broadcasting.

    DCHECK(grad_outputs[0]);
    grad_inputs[0] = grad_outputs[0];
    grad_inputs[1] = grad_outputs[0];

    grad_inputs[0]->Ref();
    grad_inputs[1]->Ref();
    return absl::OkStatus();
  }

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

  void RunGraphsAndCompareOutputs(TF_Output* grad_outputs,
                                  TF_Output* expected_grad_outputs) {
    std::unique_ptr<CSession> csession(new CSession(graph_, s_));
    std::unique_ptr<CSession> expected_csession(
        new CSession(expected_graph_, s_));

    std::vector<TF_Output> grad_outputs_vec;
    grad_outputs_vec.assign(grad_outputs, grad_outputs + 2);
    csession->SetOutputs(grad_outputs_vec);

#include "tensorflow/c/eager/gradients.h"

namespace tensorflow {
namespace gradients {
// Ignores `grad_outputs` and sets all entries in grad_inputs to nullptr.
class NotDifferentiableGradientFunction : public GradientFunction {
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override;
};

//                  absl::Span<AbstractTensorHandle* const> grad_inputs,
//                  absl::Span<AbstractTensorHandle*> grad_outputs) override {
//     grad_outputs[0] = grad_inputs[0];
//     grad_outputs[1] = grad_inputs[0];
//     grad_outputs[0]->Ref();
//     grad_outputs[1]->Ref();
//     return OkStatus();
//   }
//   ~AddGradientFunction() override {}
// };
//

 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    CHECK_EQ(grad_outputs.size(), 1);
    CHECK_EQ(grad_inputs.size(), 1);
    grad_inputs[0] = grad_outputs[0];
    if (grad_inputs[0]) {
      grad_inputs[0]->Ref();
    }
    return absl::OkStatus();
  }
};

#include "tensorflow/c/experimental/gradients/not_differentiable.h"

namespace tensorflow {
namespace gradients {
Status NotDifferentiableGradientFunction::Compute(
    AbstractContext* ctx, absl::Span<AbstractTensorHandle* const> grad_outputs,
    absl::Span<AbstractTensorHandle*> grad_inputs) {
  for (int i = 0; i < grad_inputs.size(); i++) {
    grad_inputs[i] = nullptr;
  }
  return OkStatus();
}

            n->name(),
            "). This is a bug. Please file an issue at "
            "https://github.com/tensorflow/tensorflow/issues.");
      }
    }
  }

  // Unpack the results from grad_outputs_arg.
  TFOutputsFromOutputs(dy_arg, dy);
#endif  // defined(IS_MOBILE_PLATFORM) || defined(IS_SLIM_BUILD)
}

// TF_Session functions ----------------------------------------------