}
	baseMountPath := plugin.generateBlockDeviceBaseGlobalPath()
	for _, ref := range refs {
		if mount.PathWithinBase(ref, baseMountPath) {
			// If the global mount for block device exists, the source is block
			// device.
			// The resolved device path may not be the exact same as path in
			// local PV object if symbolic link is used. However, it's the true
			// source and can be used in reconstructed volume.

// CHECK-NEXT: [[out_chain:%.*]], [[value:%.*]] = tfrt_fallback_async.executeop.seq([[in_chain]]) key(0) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.ReadVariableOp"({{.*}})
// CHECK-NEXT: [[res:%.*]] = tfrt_fallback_async.executeop key(1) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.GatherV2"([[value]], {{.*}}, [[const]])

		}
		// If the volume has device to mount, we mark its device as uncertain.
		if gvl.deviceMounter != nil || gvl.blockVolumeMapper != nil {
			deviceMountPath, err := getDeviceMountPath(gvl)
			if err != nil {
				klog.ErrorS(err, "Could not find device mount path for volume", "volumeName", gvl.volumeName)
				continue
			}

  const std::string name_;
  int plugin_id_value_;
  stream_executor::ExecutorCache executor_cache_;
};

class CEvent : public Event {
 public:
  CEvent(SP_Device* device, SP_StreamExecutor* stream_executor)
      : device_(device),
        stream_executor_(stream_executor),
        event_handle_(nullptr) {}
  ~CEvent() override { Destroy(); }

  Event::Status PollForStatus() override {

  // CHECK: tfrt_fallback_async.const_dense_tensor  dense<1.000000e+00> : tensor<1xbf16>
  %1 = "tf.Const"() {device = "/device:CPU:0", value = dense<[1.0]> : tensor<1xbf16>} : () -> tensor<4xbf16>
  // CHECK: corert.executeop({{.*}}) "tf.Const"() {dtype = ui64, value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : 1
  %2 = "tf.Const"() {device = "/device:GPU:0", value = dense<[1, 2, 3, 4]> : tensor<4xui64>} : () -> tensor<4xui64>
  func.return %0 : tensor<4xui64>

  if (cc_ctx->op_device_context() == nullptr) {  // CPU Device
    status->status = absl::FailedPreconditionError(
        "Accessing device stream is not supported for a CPU device.");
    return nullptr;
  } else if (!cc_ctx->op_device_context()->IsPluggableDevice()) {
    status->status = absl::FailedPreconditionError(
        "Accessing device stream is only supported for pluggable devices.");
    return nullptr;

  %0 = "tf.MatMul"(%arg0, %arg0) {device = "", transpose_a = false, transpose_b = false} : (tensor<10x10xi32>, tensor<10x10xi32>) -> tensor<10x10xi32>
  %1 = "tf.MatMul"(%arg0, %arg0) {device = "", transpose_a = false, transpose_b = false} : (tensor<10x10xi32>, tensor<10x10xi32>) -> tensor<10x10xi32>

// After globalMapPath is removed, fd lock by loopback for the device can
// be released safely because no one can consume the device at this point.
// At last, device open status will be checked just in case.
// If all steps are completed, the device is marked as unmounted.
func (og *operationGenerator) GenerateUnmapDeviceFunc(
	deviceToDetach AttachedVolume,

  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6",  per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %1 = "tfl.mul"(%0, %arg2) {fused_activation_function = "RELU6", per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

#include "tensorflow/compiler/tf2xla/shape_util.h"
#include "tensorflow/compiler/tf2xla/type_util.h"

namespace tensorflow {

void XlaHostSendDeviceContext::CopyCPUTensorToDevice(
    const Tensor* cpu_tensor, Device* device, Tensor* device_tensor,
    StatusCallback done, bool sync_dst_compute) const {
  auto status = stream_->Memcpy(device_memory_base_, cpu_tensor->data(),
                                device_memory_base_->size());