I can try switching to gpu-enabled machines just to see if that path can be made to work but the services queue shouldn't need gpu so I hope we figure out running the pipeline task on cpu nodes