input_tensors_vector[i] = &input_tensor;
    }
  }

  // Create an inference context with dummy values, which will be updated later.
  InferenceContext c(TF_GRAPH_DEF_VERSION, node_def, op_reg_data->op_def,
                     std::vector<ShapeHandle>(num_inputs), input_tensors_vector,
                     {},
                     std::vector<std::unique_ptr<std::vector<ShapeAndType>>>());

  // Set input_shapes.

  for (int component_index = 0; component_index < tensors_.size();
       ++component_index) {
    // TODO(allenl): Add a C API for summarizing tensors. Currently custom
    // devices limiting themselves to a C API (for ABI compatibility) would need
    // to implement summarization for component tensors themselves.
    ImmediateExecutionTensorHandle* component =
        tensorflow::unwrap(tensors_[component_index].get());

  }

  // Populate return_tensors
  DCHECK(tf_results->return_tensors.empty());
  tf_results->return_tensors.resize(results.return_tensors.size());
  for (int i = 0; i < results.return_tensors.size(); ++i) {
    tf_results->return_tensors[i].oper =
        ToOperation(results.return_tensors[i].first);
    tf_results->return_tensors[i].index = results.return_tensors[i].second;
  }

  // Populate return_nodes

  TestHandleFromDLPack(status, ctx, {4, 1, 2}, {2, 0, 1});
  TestHandleFromDLPack(status, ctx, {4, 3, 1}, {3, 1, 1});
  TestHandleFromDLPack(status, ctx, {4, 3, 1}, {3, 1, 0});
  // Test that empty tensors can have any strides.
  TestHandleFromDLPack(status, ctx, {4, 0, 2}, {0, 2, 1});
  TestHandleFromDLPack(status, ctx, {4, 0, 2}, {0, 1, 1});
  TestHandleFromDLPack(status, ctx, {4, 0, 2}, {0, 0, 1});

  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ASSERT_EQ(2, TFE_TensorDebugInfoOnDeviceNumDims(debug_info));
  // Shape is the same for CPU tensors.
  EXPECT_EQ(3, TFE_TensorDebugInfoOnDeviceDim(debug_info, 0));
  EXPECT_EQ(2, TFE_TensorDebugInfoOnDeviceDim(debug_info, 1));

  TFE_DeleteTensorDebugInfo(debug_info);
  TFE_DeleteTensorHandle(h);

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

#include "absl/types/span.h"
#include "tensorflow/c/eager/abstract_tensor_handle.h"
#include "tensorflow/c/eager/immediate_execution_context.h"
#include "tensorflow/c/eager/immediate_execution_tensor_handle.h"
#include "tensorflow/c/experimental/ops/array_ops.h"
#include "tensorflow/c/experimental/ops/math_ops.h"
#include "tensorflow/c/experimental/ops/nn_ops.h"

}
TF_AbstractTensor* TF_OutputListGet(TF_OutputList* o, int i) {
  return wrap(unwrap(o)->outputs[i]);
}
void TF_OutputListPushBack(TF_OutputList* o, TF_AbstractTensor* tensor,
                           TF_Status* s) {
  unwrap(o)->outputs.push_back(unwrap(tensor));
}

void TF_AbstractOpSetOpType(TF_AbstractOp* op, const char* const op_type,
                            TF_Status* s) {
  tsl::Set_TF_Status_from_Status(

  {
    std::unique_ptr<ParallelTensor> float_tensors =
        parallel_device.ScalarsFromSequence<float>({10.0, 11.0}, context.get(),
                                                   status.get());
    ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
    ExpectScalarEq<float>(float_tensors->tensor(0), 10.0);
    ExpectScalarEq<float>(float_tensors->tensor(1), 11.0);
  }

  {

      numerical_grad.reset(numerical_grad_raw);
    }

    TF_Tensor* numerical_tensor;
    s = GetValue(numerical_grad.get(), &numerical_tensor);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    auto num_elem_numerical = TF_TensorElementCount(numerical_tensor);

    TF_Tensor* analytical_tensor;
    s = GetValue(outputs[i], &analytical_tensor);
    ASSERT_EQ(errors::OK, s.code()) << s.message();

  ASSERT_EQ(errors::OK, s.code()) << s.message();

  TF_Tensor* result_tensor;
  s = GetValue(outputs[0], &result_tensor);
  ASSERT_EQ(errors::OK, s.code()) << s.message();
  auto result_value = static_cast<float*>(TF_TensorData(result_tensor));
  EXPECT_EQ(*result_value, 1.0);
  outputs[0]->Unref();
  TF_DeleteTensor(result_tensor);
  result_tensor = nullptr;
}

INSTANTIATE_TEST_SUITE_P(