// TODO(bfontain): Check for other attributes.
      replicated_output->setAttr(kDevice, builder.getStringAttr(""));
      TF::TPUReplicatedInputOp input = builder.create<TF::TPUReplicatedInputOp>(
          op->getLoc(), result.getType(), replicated_output.getResults());
      input->setAttr(kDevice, builder.getStringAttr(""));
      mlir::Value new_value = input.getOutput();

    }
    auto tag = tensorflow::strings::StrCat("Device #", i, " (", name, ")");
    // Copy to device
    TFE_TensorHandle* hdevice =
        TFE_TensorHandleCopyToDevice(hcpu, ctx, name.c_str(), status.get());
    if (TF_GetCode(status.get()) != TF_OK) {
      ADD_FAILURE() << tag << " -- " << TF_Message(status.get());
      continue;
    }
    // Copy from device to the same device.
    TFE_TensorHandle* hdevice2 =

  %outputs_8 =  "tf.Const"() {device = "", value = dense<1> : tensor<i32>} : () -> tensor<i32>
  %outputs_10 =  "tf.Const"() {device = "", value = dense<256> : tensor<i32>} : () -> tensor<i32>
  %outputs_12 =  "tf.Const"() {device = "", value = dense<0> : tensor<i32>} : () -> tensor<i32>

    %cst_4 = "tf.Const"() {device = "", value = dense<0.587548196> : tensor<f32>} : () -> tensor<f32>
    %cst_5 = "tf.Const"() {device = "", value = dense<0.0235294122> : tensor<f32>} : () -> tensor<f32>
    %0 = "tf.Div"(%arg0, %cst_4) {device = ""} : (tensor<?x?x?x3xf32>, tensor<f32>) -> tensor<?x?x?x3xf32>

  return
}
```
### `-tf-device-assignment-by-func-attr`

_Device assignment in TF dialect using the device specified in the function attribute._

### `-tf-device-cluster-formation`

_Form clusters from instructions assigned to same device_

Clusters operations with the same device assignment id. For each

  %1 = "tfl.mul"(%0, %arg2) {tac.device = "GPU", fused_activation_function = "RELU6", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %2 = "tfl.add"(%arg0, %arg3) {tac.device = "GPU", fused_activation_function = "RELU6", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "replicate", device = "/device:TPU:0", num_replicas = 1, topology = "topology"} : () -> ()
  func.return
}

// CHECK:      "tf.opA"() ({
// CHECK-NEXT: "tf.opB"
// CHECK-NOT: _replication_info = "replicate"
// CHECK-NOT:  device = "/device:TPU:0"

	DICS_ENABLE     DICS_STATE = 0x00000001 // The device is being enabled.
	DICS_DISABLE    DICS_STATE = 0x00000002 // The device is being disabled.
	DICS_PROPCHANGE DICS_STATE = 0x00000003 // The properties of the device have changed.
	DICS_START      DICS_STATE = 0x00000004 // The device is being started (if the request is for the currently active hardware profile).

module {
  func.func @conv_with_int_per_layer(%arg0: tensor<*xf32>) -> tensor<*xf32> {
    %cst = "tf.Const"() {device = "", value = dense<[-0.0192535277, -5.998660e-03]> : tensor<2xf32>} : () -> tensor<2xf32>

	require.NoError(t, err)
	defer os.RemoveAll(socketDir)
	m, _, p, _ := setupInProbeMode(t, []*pluginapi.Device{}, func(n string, d []pluginapi.Device) {}, socketName, pluginSocketName)
	cleanup(t, m, p)
}

// Tests that the device plugin manager correctly handles registration and re-registration by
// making sure that after registration, devices are correctly updated and if a re-registration