TF_ASSIGN_OR_RETURN(auto device_context, GetDeviceContextDefault());
  device_context->Ref();
  *out_context = device_context;
  return absl::OkStatus();
}

// Warn about XLA_CPU/XLA_GPU exactly once.
static void ShowXlaDeviceDeprecationWarning(
    absl::string_view compilation_device_name) {
  static absl::once_flag once;
  if (absl::StrContains(compilation_device_name, "CPU") ||

  // TODO(cheshire): Avoid duplication with framework/op_kernel.h
  DeviceContext* device_context = nullptr;
  if (device != nullptr) {
    TF_RETURN_IF_ERROR(device->TryGetDeviceContext(&device_context));
    bool using_default_context = false;
    auto cleanup = absl::MakeCleanup([&] {
      if (device_context != nullptr && !using_default_context) {
        device_context->Unref();
      }
    });
    if (device_context == nullptr) {

  Status TryGetDeviceContext(DeviceContext** out_context) override
      TF_LOCKS_EXCLUDED(mu_);

  Status MakeTensorFromProto(const TensorProto& tensor_proto,
                             const AllocatorAttributes alloc_attrs,
                             Tensor* tensor) override TF_LOCKS_EXCLUDED(mu_);

  Status MakeTensorFromProto(DeviceContext* device_context,

            tsl::MakeConstructedAsyncValueRef<std::unique_ptr<se::Event>>(
                std::move(event));

        Rendezvous::Args args;
        // Rendezvous::Args owns the device context pointer.
        args.device_context = new XlaHostRecvDeviceContext(
            stream, device_memory_base, shape, done_event);

        Tensor host_tensor;
        TF_RETURN_IF_ERROR(

    SetDevice(device_type, std::move(device));

    XlaShapeLayoutHelpers::ShapeDeterminationFns shape_fns{
        UseNoPreferenceLayoutFn(), IdentityShapeRepresentationFn()};
    device_context_ = core::RefCountPtr<DeviceContext>(
        new PjRtDeviceContext(shape_fns, /*use_pjrt_tensor_buffer=*/true));

    // Get the host allocator.
    AllocatorAttributes host_alloc_attr;
    host_alloc_attr.set_on_host(true);

        device_type,
        xla::GetTfrtCpuClient(/*asynchronous=*/true, /*cpu_device_count=*/1)
            .value()));

    // device_context_ should be a PjRtDeviceContext.
    TF_CHECK_OK(device_->TryGetDeviceContext(&device_context_));

    // Get the host allocator.
    AllocatorAttributes host_alloc_attr;
    host_alloc_attr.set_on_host(true);

      tensorflow::AsyncValueTensor::FromTensor(output_tensor);
  result_tensor->GetBuffer()->GetReadyFuture().OnReady(std::move(done));
}

void PjRtDeviceToDeviceCopy(DeviceContext* send_dev_context,
                            DeviceContext* recv_dev_context, Device* src,
                            Device* dst, AllocatorAttributes src_alloc_attr,
                            AllocatorAttributes dst_alloc_attr,

// strictly hierarchical, that is, transfers between devices on different hosts
// can only take place using the host network.
void TpuDeviceToDeviceCopy(DeviceContext* src_dev_context,
                           DeviceContext* dst_dev_context, Device* src,
                           Device* dst, AllocatorAttributes src_allocator_attrs,
                           AllocatorAttributes dst_allocator_attrs,