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) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "CollectiveReduce", status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);