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();

  TFE_TensorHandleGetStatus(components[0].get(), status);
  if (!status->status.ok()) {
    return nullptr;
  }

  TF_DataType dtype = TFE_TensorHandleDataType(components[0].get());
  // Verify that the combined TensorHandle's dtype matches all of the component
  // dtypes.
  for (TensorHandlePtr& component : components) {
    TFE_TensorHandleGetStatus(component.get(), status);
    if (!status->status.ok()) {
      return nullptr;

  components[0] = negative_one_cpu.get();
  components[1] = negative_one_cpu.get();
  TensorHandlePtr first_negative_one = CreatePerDeviceValues(
      context.get(), components, first_device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  components[0] = first_negative_one.get();
  components[1] = negative_one_cpu.get();

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

    ExpectScalarEq<float>(components[0].get(), 20.);
    ExpectScalarEq<float>(components[1].get(), 20.);

    std::string first_device =
        TFE_TensorHandleBackingDeviceName(components[0].get(), status.get());

 public:
  // Construct a ParallelTensor from TensorHandles placed on the component
  // devices of a ParallelDevice. If called, ParallelTensor::Shape inspects
  // `components` to determine a shape.
  static std::unique_ptr<ParallelTensor> FromTensorHandles(
      const ParallelDevice& parallel_device,
      std::vector<TensorHandlePtr> components, TF_Status* status);

// two component devices, running `x = TPUReplicatedInput(inputs=[a, b])` on the
// parallel device creates a parallel tensor `x` with `a` on the first of
// `underlying_devices` and `b` on the second. Running `a_unpacked, b_unpacked =
// TPUReplicatedOutput(input=x, num_replicas=2)` un-packs the parallel tensor
// into its components.
//

    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);

        "--quiet", "--wait", "--nocache", \
        "--add", "Microsoft.VisualStudio.Workload.VCTools", \
        "--add", "Microsoft.VisualStudio.Component.VC.Tools.x86.x64", \
        "--add", "Microsoft.VisualStudio.Component.Windows10SDK.19041" -Wait; \
    Start-Process -FilePath C:/TEMP/vs_buildtools.exe -ArgumentList "--installPath", "C:/VS", \
        "--quiet", "--wait", "--nocache", "--includeOptional", \

          " inputs to TPUReplicatedInput, but got ", inputs.size()));
      TF_SetStatus(status, TF_INVALID_ARGUMENT, message.c_str());
      return result;
    }
    std::vector<TensorHandlePtr> components;
    components.reserve(inputs.size());
    for (int i = 0; i < inputs.size(); ++i) {
      if (absl::holds_alternative<ParallelTensor*>(inputs[i])) {
        std::string message(absl::StrCat(

    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",
    ],