&FloatDeallocator, nullptr),
      TF_DeleteTensor);
  return TensorHandlePtr(TFE_NewTensorHandle(tensor.get(), status));
}

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>
void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,

#include "tensorflow/c/eager/parallel_device/parallel_device_lib.h"
#include "tensorflow/c/eager/tfe_tensorhandle_internal.h"
#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); }
};

  // finished, if any.
  //
  // `cancellation_manager` must live until after `Join` finishes and pending
  // `is_async` operations finish. In addition to allowing the caller to cancel
  // the operation, its `StartCancel` method will be called if op execution
  // fails on any device in order to cancel the others.
  void StartExecute(TFE_Context* context, const char* operation_name,