#       export TF_ANY_TARGETS="quoted list of targets, like on the command line"
#       export TF_ANY_MODE="test" or "build" or "run" (default: "test")
#       ./any.sh
#
# 2. RUN ANY OTHER SCRIPT AND ENV WITH NO SIDE EFFECTS (NO UPLOADS)
#    To use:
#       export TFCI=ci/official/envs/env_goes_here
#       export TF_ANY_SCRIPT=ci/official/wheel.sh
#       ./any.sh
#
# 3. DO THE SAME WITH A LOCAL CACHE OR RBE:

            CACHEBUSTER=${{ steps.vars.outputs.DATE }}
          tags: |
            tensorflow/build:${{ steps.vars.outputs.REF }}-${{ matrix.python-version }}
            gcr.io/tensorflow-sigs/build:${{ steps.vars.outputs.REF }}-${{ matrix.python-version }}
          cache-from: type=registry,ref=tensorflow/build:${{ steps.vars.outputs.REF }}-${{ matrix.python-version }}
          cache-to: type=inline
      -

// once (i.e., on the main function side) in running distributed functions.
// This test creates a cluster with two workers, create a variable on the
// second worker, and run a distributed function (VariableAddFunction) whose ops
// span the local and remote workers. If the graph optimization pass is executed
// on both the main function side and the component function side, an error will

std::unique_ptr<ParallelTensor> ParallelDevice::DeviceIDs(
    TFE_Context* context, TF_Status* status) const {
  std::vector<int32_t> ids;
  ids.reserve(num_underlying_devices());
  for (int i = 0; i < num_underlying_devices(); ++i) {
    ids.push_back(i);
  }
  return ScalarsFromSequence<int32_t>(ids, context, status);
}

absl::optional<std::vector<std::unique_ptr<ParallelTensor>>>
ParallelDevice::Execute(TFE_Context* context,