auto device = op->getAttrOfType<StringAttr>(kDevice);
      for (auto *operand : operands)
        propagateIfChanged(operand, operand->SetDevice(device));
    } else {
      // Propagate device through other ops. These ops might have their
      // own device annotation, but that's fine. We only care about
      // where the TPUExecute ops live.
      StringAttr device;
      for (const Device *d : results) {

  // TODO(b/177376459): Update if needed to make testing easy.
  if (!module_) {
    for (const auto& device : device_specs) {
      auto* hardware = this->GetTargetHardware(device);
      if (hardware == nullptr) continue;
      if (hardware->IsOpSupported(op)) {
        SetAnnotation(op, kDevice, device, builder);
        device_is_set = true;
        break;
      }
    }
  } else {

  return name;
}

// Get the target annotation form the op.
inline std::optional<std::string> GetTargetAnnotation(Operation* op) {
  auto device = op->getAttrOfType<StringAttr>(kDevice);
  if (device == nullptr || device.getValue().empty()) return std::nullopt;

  return GetCanonicalHardwareName(device.getValue().str());
}

// Get inference type attribute from the operation if available.

// `{ tac.device = "GPU", tac.inference_type = "FLOAT"}` to a function
// with the matching attributes. Assumed is that device type "CPU"
// is the only device that is allowed to call other devices. I.e. ancestors of a
// "CPU" `Operation` may only `Operations` without a device or other "CPU"
// `Operations`. Implied is that "CPU" ops may contain subgraphs of different

  for (const auto& device : devices) {
    if (inference_type == QUANTIZED_INT8) {
      all_device_inference_types.push_back({device, QUANTIZED_INT8});
    } else if (inference_type == QUANTIZED_UINT8) {
      all_device_inference_types.push_back({device, QUANTIZED_UINT8});
    }

    // We will alway enable float.
    all_device_inference_types.push_back({device, FLOAT});
  }

      // TODO(bfontain): Check for other attributes.
      replicated_output->setAttr(kDevice, builder.getStringAttr(""));
      TF::TPUReplicatedInputOp input = builder.create<TF::TPUReplicatedInputOp>(
          op->getLoc(), result.getType(), replicated_output.getResults());
      input->setAttr(kDevice, builder.getStringAttr(""));
      mlir::Value new_value = input.getOutput();

    return std::nullopt;

  if (!HasValidHardwareTarget(op)) return std::nullopt;

  auto device = op->getAttrOfType<mlir::StringAttr>(mlir::TFL::tac::kDevice);
  if (device == nullptr) return std::nullopt;

  llvm::StringRef device_name_str = device.getValue();
  return device_name_str.str();
}

std::optional<std::vector<float>> GetPerDeviceCosts(

      // TODO(bfontain): Check for other attributes.
      replicated_output->setAttr(kDevice, builder.getStringAttr(""));
      TF::TPUReplicatedInputOp input = builder.create<TF::TPUReplicatedInputOp>(
          op->getLoc(), result.getType(), replicated_output.getResults());
      input->setAttr(kDevice, builder.getStringAttr(""));
      mlir::Value new_value = input.getOutput();

          // Set interface_name & target to the call_op as well.
          new_call->setAttr(kInterfaceNameAttr,
                            builder->getStringAttr(interface_name));
          new_call->setAttr(
              kDevice,
              builder->getStringAttr(preferred_inference_device_type.hardware));
          new_call->setAttr(
              kInferenceType,
              builder->getStringAttr(GetInferenceString(

// RUN: tac-translate -input-mlir -output-mlir -device-specs=NNAPI %s -o - 2>&1 | FileCheck %s

module {
  // CHECK-LABEL: main
  func.func @main(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
    %0 = "tfl.squared_difference"(%arg0, %arg1) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
    func.return %0 : tensor<4xf32>
    // CHECK:  [[VAL_0:%.*]] = tfl.sub %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<4xf32>