We should add some level of CI testing at sf1 (which should work on a single gpu) to catch breakages earlier. Came up as a discussion from #244 .
