}
    auto begin = data.begin();
    if (offset > pos) {
      // The block begins before the slice we're reading.
      begin += offset - pos;
    }
    auto end = data.end();
    if (pos + data.size() > offset + n) {
      // The block extends past the end of the slice we're reading.
      end -= (pos + data.size()) - (offset + n);
    }
    if (begin < end) {
      size_t bytes_to_copy = end - begin;

  explicit DeviceThread(const std::string& device, const bool is_async,
                        const int in_flight_nodes_limit)
      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in
        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //

    tensorflow::ImmediateExecutionTensorHandle* unwrapped_handle =
        tensorflow::unwrap(handles[i]);
    if (tensorflow::CustomDeviceTensorHandle::classof(unwrapped_handle)) {
      // One of the inputs we're trying to pack is on a custom device. We'll let
      // the first custom device we see handle all of the packing.
      auto* custom_device_handle =
          tensorflow::down_cast<tensorflow::CustomDeviceTensorHandle*>(