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);
  const std::string merge_op("Add");
  TFE_OpSetAttrString(op.get(), "merge_op", merge_op.c_str(),

// Return a shape op fetching the shape of `a`.
TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return an allreduce op adding up input tensor `in` from `group_size` workers.
TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size);

// Return a SendOp op `op_name` with send input tensor `in` and attributes
// `send_device`, `recv_device`, and `send_device_incarnation` set.

  TF_DeleteStatus(status);

  TFE_OpSetAttrType(op, "T", TFE_TensorHandleDataType(in));
  TFE_OpSetAttrInt(op, "group_size", group_size);
  TFE_OpSetAttrInt(op, "group_key", 123);
  TFE_OpSetAttrInt(op, "instance_key", 456);
  TFE_OpSetAttrString(op, "merge_op", "Add", 3);
  TFE_OpSetAttrString(op, "final_op", "Id", 2);
  std::vector<int64_t> subdiv_offsets;

           {"condition", "x"},
           {{"T", std::vector<DataType>{DT_FLOAT}}}},
          {{"y"},
           "CollectiveReduce",
           {"x"},
           {{"T", DT_FLOAT},
            {"group_size", static_cast<int>(devices.size())},
            {"group_key", 0},
            {"instance_key", 0},
            {"merge_op", "Add"},
            {"final_op", "Id"},
            {"subdiv_offsets", std::vector<int>()}},