`XNNPACK` delegate automatically when the model has a `fp32` operation.
*   GPU
    *   Allow GPU acceleration starting with internal graph nodes
    *   Experimental support for quantized models with the Android GPU delegate
    *   Add GPU delegate whitelist.
    *   Rename GPU whitelist -> compatibility (list).
    *   Improve GPU compatibility list entries from crash reports.
*   NNAPI
    *   Set default value for

### Hardware attacks

Physical GPUs or TPUs can also be the target of attacks. [Published
research](https://scholar.google.com/scholar?q=gpu+side+channel) shows that it
might be possible to use side channel attacks on the GPU to leak data from other
running models or processes in the same system. GPUs can also have
implementation bugs that might allow attackers to leave malicious code running

  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).

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

# limitations under the License.
# ==============================================================================
source "${BASH_SOURCE%/*}/utilities/setup.sh"

# Record GPU count and CUDA version status
if [[ "$TFCI_NVIDIA_SMI_ENABLE" == 1 ]]; then
  tfrun nvidia-smi
fi

# Update the version numbers for Nightly only
if [[ "$TFCI_NIGHTLY_UPDATE_VERSION_ENABLE" == 1 ]]; then

  if [[ "$TFCI_PYTHON_VERSION" == "3.13" ]]; then
    "$python" -m pip install numpy==1.26.4
  else
    "$python" -m pip install numpy==1.26.0
  fi
fi
if [[ "$TFCI_BAZEL_COMMON_ARGS" =~ gpu|cuda ]]; then
  echo "Checking to make sure tensorflow[and-cuda] is installable..."
  "$python" -m pip install "$(echo *.whl)[and-cuda]" $TFCI_PYTHON_VERIFY_PIP_INSTALL_ARGS
else

        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

А затем сразу после `yield` мы выгружаем модель. Этот код будет выполнен после того, как приложение закончит обрабатывать запросы, непосредственно перед shutdown. Это может, например, освободить ресурсы, такие как память или GPU.

/// tip | Совет

`shutdown` произойдёт, когда вы останавливаете приложение.

Возможно, вам нужно запустить новую версию, или вы просто устали от него. 🤷

///

### Функция lifespan { #lifespan-function }

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. 🤷

///

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

/// tip | Dica

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

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

///