(environ_cp.get('TF_NEED_ROCM', None) == '1')):
      test_and_build_filters += ['-no_windows_gpu', '-no_gpu']
    else:
      test_and_build_filters.append('-gpu')
  elif is_macos():
    test_and_build_filters += ['-gpu', '-nomac', '-no_mac', '-mac_excluded']
  elif is_linux():
    if ((environ_cp.get('TF_NEED_CUDA', None) == '1') or
        (environ_cp.get('TF_NEED_ROCM', None) == '1')):

  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  ASSERT_EQ(0, device_id) << device_id;

  // Disable the test if no GPU is present.
  string gpu_device_name;
  if (GetDeviceName(ctx, &gpu_device_name, "GPU")) {
    TFE_TensorHandle* hgpu = TFE_TensorHandleCopyToDevice(
        hcpu, ctx, gpu_device_name.c_str(), status.get());

    # Googlers: search for "test_tf_whl_size"
    case "$TF_WHEEL" in
        # CPU:
        *cpu*manylinux*) LARGEST_OK_SIZE=220 ;;
        # GPU:
        *manylinux*)     LARGEST_OK_SIZE=580 ;;
        # Unknown:
        *)
            echo "The wheel's name is in an unknown format."
            exit 1
            ;;
    esac

You may invoke a CI script of your choice by following these instructions:

```bash
cd tensorflow-git-dir

# Here is a single-line example of running a script on Linux to build the
# GPU version of TensorFlow for Python 3.12, using the public TF bazel cache and
# a local build cache:
TFCI=py312,linux_x86_cuda,public_cache,disk_cache ci/official/wheel.sh

TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,
              TFE_TensorHandle* axis);

// If there is a device of type `device_type`, returns true
// and sets 'device_name' accordingly.
// `device_type` must be either "GPU" or "TPU".
bool GetDeviceName(TFE_Context* ctx, tensorflow::string* device_name,
                   const char* device_type);

// Create a ServerDef with the given `job_name` and add `num_tasks` tasks in it.

  // logic to refer to the specific device chosen.
  //
  // Example: If one calls `op->SetDeviceName("/device:GPU")`, the value
  // returned by DeviceName should be "/device:GPU:*" until a particular GPU is
  // chosen for the operation by the device placement logic in the
  // executor. After that, the value returned by DeviceName will be a full
  // device name such as "/job:localhost/replica:0/task:0/device:GPU:1".

        with:
          api-level: ${{ matrix.api-level }}
          arch: ${{ matrix.api-level == '34' && 'x86_64' || 'x86' }}
          force-avd-creation: false
          emulator-options: -no-window -gpu swiftshader_indirect -noaudio -no-boot-anim -camera-back none
          disable-animations: false
          script: echo "Generated AVD snapshot for caching."

      - name: Run Tests

E então, logo após o `yield`, descarregaremos o modelo. Esse código será executado **após** a aplicação **terminar de lidar com as requisições**, pouco antes do *encerramento*. Isso poderia, por exemplo, liberar recursos como memória ou GPU.

/// tip | "Dica"

O `shutdown` aconteceria quando você estivesse **encerrando** a aplicação.

Talvez você precise inicializar uma nova versão, ou apenas cansou de executá-la. 🤷

///

### Função _lifespan_

And then, right after the `yield`, we unload the model. This code will be executed **after** the application **finishes handling requests**, right before the *shutdown*. This could, for example, release resources like memory or a GPU.

/// tip

The `shutdown` would happen when you are **stopping** the application.

Maybe you need to start a new version, or you just got tired of running it. 🤷

///

### Lifespan function

// Sets XLA's auto jit mode according to the specified string, which is parsed
// as if passed in XLA_FLAGS. This has global effect.
TF_CAPI_EXPORT void TF_SetXlaAutoJitMode(const char* mode);

// Returns whether the single GPU or general XLA auto jit optimizations are
// enabled through MarkForCompilationPassFlags.
TF_CAPI_EXPORT unsigned char TF_GetXlaAutoJitEnabled();

// Sets XLA's minimum cluster size. This has global effect.