INFRA Signal 163
Run Ray on TPU, Part 1: The foundations
↗
Engineers can now deploy to TPUs using official pre-built images instead of custom containers required by the previous experimental path. The integration handles the strict networking requirement of TPU slices—where workers must share a high-speed interconnect or the job hangs—by using GKE labels to schedule entire slices atomically. This eliminates the need for manual placement logic when targeting Google's accelerators.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Ray 2.55 elevates Google Cloud TPUs to a first-class accelerator with official release pipelines and pre-built images.
TPU chips are wired into fixed slices with a dedicated interconnect, requiring all workers of a multi-host model to land on the same slice to avoid hanging.
A GKE Ray TPU webhook stamps hosts with slice labels so Ray Core can reserve whole slices based on topology rather than individual chip counts.
THE CLUSTER