class OpDeleter {
 public:
  void operator()(TFE_Op* to_delete) const { TFE_DeleteOp(to_delete); }
};

using OpPtr = std::unique_ptr<TFE_Op, OpDeleter>;

class StatusDeleter {
 public:
  void operator()(TF_Status* to_delete) const { TF_DeleteStatus(to_delete); }
};

using StatusPtr = std::unique_ptr<TF_Status, StatusDeleter>;

class ExecutorDeleter {
 public:
  void operator()(TFE_Executor* to_delete) const {

#include "tensorflow/c/tf_status.h"
#include "tensorflow/c/tf_status_helper.h"

namespace tensorflow {
namespace parallel_device {
namespace {

class OpDeleter {
 public:
  void operator()(TFE_Op* to_delete) const { TFE_DeleteOp(to_delete); }
};

using OpPtr = std::unique_ptr<TFE_Op, OpDeleter>;

using MaybeParallelTensorOwned =
    absl::variant<std::unique_ptr<ParallelTensor>, TensorHandlePtr>;

                    /*input_shapes*/ {make_shape({3, 2}), make_shape({2, 4})},
                    /*input_tensors*/ {},
                    /*expected_shape*/ make_shape({3, 4}));

  // Infer shape when second operand has unknown shape.
  CheckOutputShapes(matmul_op,
                    /*input_shapes*/ {make_shape({3, 2}), unknown_shape()},
                    /*input_tensors*/ {},

  size_t finish = block_size_ * ((offset + n) / block_size_);
  if (finish < offset + n) {
    finish += block_size_;
  }
  size_t total_bytes_transferred = 0;
  // Now iterate through the blocks, reading them one at a time.
  for (size_t pos = start; pos < finish; pos += block_size_) {
    Key key = std::make_pair(filename, pos);