VLOG(2) << "Transfer to device as literal: " << literal.ToString() << " "
            << xla_tensor->shaped_buffer().ToString();
    if (UseMultipleStreams() &&
        !transfer_manager_->CanShapedBufferBeAccessedNow(
            stream_->parent(), xla_tensor->shaped_buffer())) {
      // Initially wait for the compute stream so that memory allocations are
      // synchronized.

    // Wait for the destination tensor buffers to be ready, if they are not
    // available for an immediate write.
    if (!dst_xla_context->transfer_manager()->CanShapedBufferBeAccessedNow(
            dst_compute_stream->parent(), xla_output->shaped_buffer())) {
      TF_RETURN_IF_ERROR(
          dst_device_to_device_stream->WaitFor(dst_compute_stream));