TFE_TensorDebugInfo* debug_info, int dim_index);

// Description of the TensorFlow op to execute.
//
// Assumes that the provided 'ctx' outlives the returned TFE_Op, i.e.,
// TFE_DeleteOp() is called before TFE_DeleteContext().
//
// Very similar to TF_OperationDescription with some differences:
// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,
//     TF_AddInputList

  TFE_DeleteTensorHandle(retvals[0]);
  for (auto* h : handles_task0) {
    TFE_DeleteTensorHandle(h);
  }
  for (auto* h : handles_task2) {
    TFE_DeleteTensorHandle(h);
  }

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);

    // on to one TFE_Op per device and just call TFE_ResetOp to avoid parsing
    // device names repeatedly.
    std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> const_op(
        TFE_NewOp(context, "Const", status), TFE_DeleteOp);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    TFE_OpSetDevice(const_op.get(), underlying_devices_[device_index].c_str(),
                    status);

      TFE_OpAddInput(op, input, s);
    }
    if (TF_GetCode(s) != TF_OK) return;
  }
  std::vector<TFE_TensorHandle*> op_outputs(*num_outputs);
  TFE_Execute(op, op_outputs.data(), num_outputs, s);
  TFE_DeleteOp(op);
  if (TF_GetCode(s) != TF_OK) return;
  std::vector<TFE_TensorHandle*> unwrapped_outputs;
  unwrapped_outputs.reserve(op_outputs.size());
  for (auto* handle : op_outputs) {

  status->status = new_op->Reset(op_or_function_name, nullptr);
  if (!status->status.ok()) {
    new_op->Release();
    new_op = nullptr;
  }
  return tensorflow::wrap(new_op);
}

void TFE_DeleteOp(TFE_Op* op) {
  if (op == nullptr) {
    return;
  }

  tensorflow::unwrap(op)->Release();
}

const char* TFE_OpGetName(const TFE_Op* op, TF_Status* status) {

#include "tsl/platform/errors.h"
#include "tsl/platform/thread_annotations.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>;

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

#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>;

using MaybeParallelTensorUnowned =