// The device tensor should always be fresh.
    TF_RET_CHECK(!xla_tensor->has_shaped_buffer());

    TF_RETURN_IF_ERROR(
        xla_tensor->AllocateShapedBuffer(device_tensor->dtype(), shape, client_,
                                         stream_->parent()->device_ordinal()));

    // The cpu_tensor and literal that we created here hold the data of host

                            input->shape(), input->dtype(),
                            /*use_fast_memory=*/false, layout_preference));
    TF_RETURN_IF_ERROR(xla_output->AllocateShapedBuffer(
        input->dtype(), shape, dst_xla_context->client(), dst_device_ordinal));

    VLOG(2) << "TpuDeviceToDeviceCopy: src: "
            << src_compute_stream->parent()->device_ordinal() << ", "