// 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
    grad_inputs[1] = nullptr;
    return absl::OkStatus();
  }
  ~SparseSoftmaxCrossEntropyWithLogitsGradientFunction() override {}

 private:
  vector<AbstractTensorHandle*> forward_outputs_;
};

  // Convert from AbstractTensor to TF_Tensor.
  TF_Tensor* theta_tensor;
  TF_RETURN_IF_ERROR(GetValue(theta, &theta_tensor));

  // Get number of elements and fill data.
  int num_elems = TF_TensorElementCount(theta_tensor);
  vector<float> theta_data(num_elems);
  memcpy(theta_data.data(), TF_TensorData(theta_tensor),
         TF_TensorByteSize(theta_tensor));

                      {x.get()}, absl::MakeSpan(outputs),
                      /*use_function=*/!std::get<2>(GetParam()));
  ASSERT_EQ(error::INVALID_ARGUMENT, s.code());
  ASSERT_EQ(
      "Provided null gradient_function for 'Neg'.\nIf the intent is to treat "
      "this op as non-differentiable consider using RegisterNotDifferentiable "
      "or NotDifferentiableGradientFunction.",
      s.message());

    for (int i = 0; i < entry.slices_size(); ++i) {
      const auto& slice_proto = entry.slices(i);
      CHECK(filtered_keys
                .insert(EncodeTensorNameSlice(
                    string(v2_reader_->key()) /* full var's name */,
                    TensorSlice(slice_proto)))
                .second);
    }
  }

  // Second pass: adds the entries, ignoring the filtered keys.

    core::RefCountPtr<AbstractFunction> scoped_func;
    // Returning null tensors from a tf.function is not supported, so we keep
    // track of indices in the model's outputs are nullptr in this set.
    // The FunctionDef only outputs the non-null tensors. We later pad the
    // function op outputs to have nullptrs at the `null_indices`.
    absl::flat_hash_set<int> null_indices;
    {
      AbstractContextPtr func_ctx(BuildFunction(fn_name));