reachable and exploitable through production-grade, benign models.

### Compilation

Compiling models via the recommended entry points described in
[XLA](https://www.tensorflow.org/xla) and
[JAX](https://jax.readthedocs.io/en/latest/jax-101/02-jitting.html)
documentation should be safe, while some of the testing and debugging tools that
come with the compiler are not designed to be used with untrusted data and

    std::unique_ptr<CSession> expected_csession(
        new CSession(expected_graph_, s_));

    std::vector<TF_Output> grad_outputs_vec;
    grad_outputs_vec.assign(grad_outputs, grad_outputs + 2);
    csession->SetOutputs(grad_outputs_vec);
    csession->Run(s_);
    ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);
    TF_Tensor* out0 = csession->output_tensor(0);

dataflow graph across many machines in a cluster, and within a machine across multiple computational devices, including multicore CPUs, general purpose GPUs, and custom-designed ASICs known as Tensor Processing Units (TPUs). This architecture gives flexibility to the application developer, whereas in previous “parameter server” designs the management of shared state is built into the system, TensorFlow enables developers to experiment with novel optimizations and training algorithms. TensorFlow supports...

    throw "Error adding labels to PR";
  }
  if (assignees.length > 0) {
    const resp_assign = await github.rest.issues.addAssignees({
      issue_number: context.issue.number,
      owner: context.repo.owner,
      repo: context.repo.repo,
      assignees: assignees
    });
    if (resp_assign.status >= 400) {
      console.log(resp_assign);
      throw "Error adding assignee to PR";
    }
  }

# kernel registration - REGISTER_OP, REGISTER_KERNEL_BUILDER, etc.
#
# This setting is currently experimental. The 'v2' implementation does _not_
# correspond to a particular, finalized design; rather, it relates to
# developing one.
#
# The current aim of the 'v2' implementation is to allow 'unused' ops and
# kernels to be discarded by the linker (to the benefit of binary size).
bool_flag(

        ld_version_int = convert_version_to_int(ld_version[4])

      # Enable if 'ld' version >= 2.35
      if ld_version_int >= 2035000:
        write_to_bazelrc(
            'build --copt="-DEIGEN_ALTIVEC_ENABLE_MMA_DYNAMIC_DISPATCH=1"')

  with_xla_support = environ_cp.get('TF_ENABLE_XLA', None)
  if with_xla_support is not None:
    write_to_bazelrc('build --define=with_xla_support=%s' %

#
# setup.packages.sh: Given a list of Ubuntu packages, install them and clean up.
# Usage: setup.packages.sh <package_list.txt>
set -e

# Prevent apt install tzinfo from asking our location (assumes UTC)
export DEBIAN_FRONTEND=noninteractive

apt-get update
# Remove commented lines and blank lines
apt-get install -y --no-install-recommends $(sed -e '/^\s*#.*$/d' -e '/^\s*$/d' "$1" | sort -u)

# Usage: setup.python.sh <pyversion> <requirements.txt>

# Sets up custom apt sources for our TF images.

# Prevent apt install tzinfo from asking our location (assumes UTC)
export DEBIAN_FRONTEND=noninteractive

# Set up shared custom sources
apt-get update
apt-get install -y gnupg ca-certificates

# Deadsnakes: https://launchpad.net/~deadsnakes/+archive/ubuntu/ppa

  ASSERT_TRUE(executed);
  auto handle_cleaner = tensorflow::gtl::MakeCleanup(
      [var_handle]() { TFE_DeleteTensorHandle(var_handle); });

  // Assign to the variable, copying to the custom device.
  std::unique_ptr<TFE_TensorHandle, decltype(&TFE_DeleteTensorHandle)> one(
      TestScalarTensorHandle(context.get(), 111.f), TFE_DeleteTensorHandle);

  if (av_tensor == nullptr) {
    TF_ASSIGN_OR_RETURN(
        *tensor, MakeTensorFromPjRtBuffer(tensor->dtype(), tensor->shape(),
                                          std::move(buffer)));
  } else {
    av_tensor->SetBuffer(std::move(buffer));
  }
  return absl::OkStatus();
}

absl::StatusOr<xla::PjRtCApiClient*> GetPjRtCApiClient(
    const DeviceType& device_type) {