std::array<TensorHandlePtr, 2> out_components;
    ExtractPerDeviceValues(context.get(), multiply_result.get(),
                           &out_components, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    ExpectScalarEq<float>(out_components[0].get(), 9.);
    ExpectScalarEq<float>(out_components[1].get(), 4.);
  }

  worker_server1.release();

    std::array<TensorHandlePtr, num_replicas>* components, TF_Status* status);

// Helper to pack `num_replicas` TFE_TensorHandles into one parallel handle.
template <std::size_t num_replicas>
TensorHandlePtr CreatePerDeviceValues(
    TFE_Context* context,
    const std::array<TFE_TensorHandle*, num_replicas>& components,
    const char* device, TF_Status* status);

namespace tensorflow {
namespace parallel_device {

// Allocate a parallel device named `device_name` which forwards operations to
// `underlying_devices`, maintaining "parallel tensors" with components placed
// on each underlying device.
//
// For example if `device_name` is
//   "/job:localhost/replica:0/task:0/device:CUSTOM:0"
// and `underlying_devices` is
//   {"/job:localhost/replica:0/task:0/device:GPU:0",

    srcs = glob([
        "aarch64-none-linux-gnu/**",
        "libexec/**",
        "lib/gcc/aarch64-none-linux-gnu/**",
        "include/**",
    ]),
)

filegroup(
    name = "compiler_components",
    srcs = [
        ":ar",
        ":as",
        ":gcc",
        ":ld",
        ":nm",
        ":objcopy",
        ":objdump",
        ":strip",
    ],