}) {device = "TPU_REPLICATED_CORE_0"} : () -> tensor<i8>
        tf_device.return %2 : tensor<i8>
      }
      return
    }
    ```

    A non replicated `tf_device.cluster_func` with the model parallelism:

    ```mlir
    func @tf_tpu_rewrite(%arg0: tensor<8xi32>) -> tensor<8xi32> {

if [[ ${KUBE_INTEGRATION_TEST_MAX_CONCURRENCY} -gt 0 ]]; then
  GOMAXPROCS=${KUBE_INTEGRATION_TEST_MAX_CONCURRENCY}
  export GOMAXPROCS
  kube::log::status "Setting parallelism to ${GOMAXPROCS}"
fi

# Give integration tests longer to run by default.
KUBE_TIMEOUT=${KUBE_TIMEOUT:--timeout=600s}
LOG_LEVEL=${LOG_LEVEL:-2}
KUBE_TEST_ARGS=${KUBE_TEST_ARGS:-}
# Default glog module settings.

	"parallelism":             "Specifies the maximum desired number of pods the job should run at any given time. The actual number of pods running in steady state will be less than this number when ((.spec.completions - .status.successful) < .spec.parallelism), i.e. when the work left to do is less than max parallelism. More info: https://kubernetes.io/docs/concepts/workloads/controllers/jobs-run-to-completion/",...

          // transitive closure of successors.  We may revisit this when
          // we see more cases needing cluster scoping in the future.
          /*edge_filter=*/nullptr);
}

// This preserves the parallelism between pipeline stages.  For example, below
// is a typical pattern of input pipelining in Tensorflow and this heuristic
// ensures Node_X and Node_Y are put into different clusters.  Without the

}

// DeepCopyInto is an autogenerated deepcopy function, copying the receiver, writing into out. in must be non-nil.
func (in *JobSpec) DeepCopyInto(out *JobSpec) {
	*out = *in
	if in.Parallelism != nil {
		in, out := &in.Parallelism, &out.Parallelism
		*out = new(int32)
		**out = **in
	}
	if in.Completions != nil {
		in, out := &in.Completions, &out.Completions
		*out = new(int32)
		**out = **in
	}

  // be less than this number when ((.spec.completions - .status.successful) < .spec.parallelism),
  // i.e. when the work left to do is less than max parallelism.
  // More info: https://kubernetes.io/docs/concepts/workloads/controllers/jobs-run-to-completion/
  // +optional
  optional int32 parallelism = 1;

  // Specifies the desired number of successfully finished pods the

}

// DeepCopyInto is an autogenerated deepcopy function, copying the receiver, writing into out. in must be non-nil.
func (in *JobSpec) DeepCopyInto(out *JobSpec) {
	*out = *in
	if in.Parallelism != nil {
		in, out := &in.Parallelism, &out.Parallelism
		*out = new(int32)
		**out = **in
	}
	if in.Completions != nil {
		in, out := &in.Completions, &out.Completions
		*out = new(int32)
		**out = **in
	}

	// for example (1) during validation to check for usage of removed plugins,
	// (2) writing config to a file, (3) initialising the scheduler.
	metav1.TypeMeta

	// Parallelism defines the amount of parallelism in algorithms for scheduling a Pods. Must be greater than 0. Defaults to 16
	Parallelism int32

	// LeaderElection defines the configuration of leader election client.
	LeaderElection componentbaseconfig.LeaderElectionConfiguration

	return autoConvert_batch_JobList_To_v1_JobList(in, out, s)
}

func autoConvert_v1_JobSpec_To_batch_JobSpec(in *v1.JobSpec, out *batch.JobSpec, s conversion.Scope) error {
	out.Parallelism = (*int32)(unsafe.Pointer(in.Parallelism))
	out.Completions = (*int32)(unsafe.Pointer(in.Completions))
	out.ActiveDeadlineSeconds = (*int64)(unsafe.Pointer(in.ActiveDeadlineSeconds))

1. Run one suite per command (e.g. `go test ./tests/integration/mysuite/...`)
1. Disable parallelism with `-p 1` (e.g. `go test -p 1 ./...`). A major disadvantage to doing this is that it will also disable
parallelism within the suite, even when explicitly specified via [RunParallel](#parallel-tests).

### Test Selection