Run Ray's AI libraries on TPUs: one YAML field keeps your whole slice together
TPU chips are wired into fixed slices, so a tensor-parallel model that spans multiple hosts has to land on one whole slice or its workers hang forever on their first collective. Google's guide walks Ray Serve, Data, and Train through that constraint, and highlights the single `topology` field that keeps a slice together — skip it and you get a deployment stuck in DEPLOYING while burning TPU-hours. Part 2 of a two-part series on Ray's higher-level libraries on GKE.