Run Ray on TPU, Part 2: Ray AI libraries

Part 2 of a guide on running Ray on Google Cloud TPUs covers the three main AI libraries: Ray Serve for LLM inference (including multi-host tensor-parallel models via the topology field), Ray Data's iter_jax_batches() for device-sharded JAX input pipelines, and JaxTrainer for distributed JAX training with checkpointing and fault tolerance. Key gotchas include importing JAX inside the worker function, using topology instead of chip counts, and setting accelerator_config.topology in Serve to avoid silent multi-host deployment failures. Official rayproject/ray:-tpu Docker images and TPU utilization metrics in the Ray Dashboard are also now available.