std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> op(
      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_OpSetAttrType(op.get(), "dtype", TF_FLOAT);
  TFE_OpSetAttrShape(op.get(), "shape", {}, 0, status.get());
  TFE_OpSetAttrString(op.get(), "container", "", 0);
  TFE_OpSetAttrString(op.get(), "shared_name", "", 0);

  TFE_OpAddInput(op, axis, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpSetAttrBool(op, "keep_dims", 1);
  TFE_OpSetAttrType(op, "Tidx", TF_INT32);
  TF_DeleteStatus(status);
  TFE_OpSetAttrType(op, "T", TFE_TensorHandleDataType(input));

  return op;
}

TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size) {

  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> handle_op(
      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_OpSetAttrType(handle_op.get(), "dtype", TF_FLOAT);
  TFE_OpSetAttrShape(handle_op.get(), "shape", /*dims=*/nullptr, /*num_dims=*/0,
                     status.get());

  TFE_Op* reshape_op = TFE_NewOp(tfe_context_, "Reshape", status_);
  CHECK_EQ(TF_OK, TF_GetCode(status_)) << TF_Message(status_);
  TFE_OpSetAttrType(reshape_op, "T", TF_FLOAT);
  TFE_OpSetAttrType(reshape_op, "Tshape", TF_INT32);
  CheckOutputShapes(reshape_op,
                    /* input_shapes*/ {unknown_shape(), unknown_shape()},

  TFE_DeleteContextOptions(opts);

  TFE_Op* var_op = TFE_NewOp(ctx, "VarHandleOp", status);
  TFE_OpSetAttrType(var_op, "dtype", TF_INT64);
  TFE_OpSetAttrShape(var_op, "shape", {}, 0, status);
  const TFE_OpAttrs* attributes = TFE_OpGetAttrs(var_op);

  TFE_Op* copy_op = TFE_NewOp(ctx, "VarHandleOp", status);
  TFE_OpSetAttrType(copy_op, "dtype", TF_FLOAT);
  TFE_OpAddAttrs(copy_op, attributes);
  unsigned char is_list = 0;

TF_CAPI_EXPORT extern void TFE_OpSetAttrBool(TFE_Op* op, const char* attr_name,
                                             unsigned char value);
TF_CAPI_EXPORT extern void TFE_OpSetAttrType(TFE_Op* op, const char* attr_name,
                                             TF_DataType value);
// If the number of dimensions is unknown, `num_dims` must be set to

                    status);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    TFE_OpSetAttrTensor(const_op.get(), "value", tensor.get(), status);
    if (TF_GetCode(status) != TF_OK) return nullptr;
    TFE_OpSetAttrType(const_op.get(), "dtype", datatype_enum);
    TFE_TensorHandle* device_handle;
    int num_outputs = 1;
    TFE_Execute(const_op.get(), &device_handle, &num_outputs, status);

                                               (value == 0) ? false : true);
  if (!s.ok()) {
    LOG(WARNING) << "Unable to set attribute: " << attr_name;
  }
}

void TFE_OpSetAttrType(TFE_Op* op, const char* attr_name, TF_DataType value) {
  auto s = tensorflow::unwrap(op)->SetAttrType(
      attr_name, static_cast<tensorflow::DataType>(value));
  if (!s.ok()) {

                  TFE_TensorHandle** out_value) {
  TF_Status* status = TF_NewStatus();
  TFE_Op* op = TFE_NewOp(ctx, "ReadVariableOp", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_OpSetAttrType(op, "dtype", TF_FLOAT);
  TFE_OpAddInput(op, var, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  int num_retvals = 1;
  TFE_Execute(op, out_value, &num_retvals, status);

  const char* device = TFE_TensorHandleDeviceName(input, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetDevice(op.get(), device, status);
  if (TF_GetCode(status) != TF_OK) return nullptr;
  TFE_OpSetAttrType(op.get(), "T", TFE_TensorHandleDataType(input));
  TFE_OpSetAttrInt(op.get(), "group_size", group_size);
  TFE_OpSetAttrInt(op.get(), "group_key", 0);
  TFE_OpSetAttrInt(op.get(), "instance_key", 0);