AWS ships no-build PyTorch fine-tune sample on Ray Train DLC

The sample is built around a division of labour: AWS assembles and tests the software stack as one unit, KubeRay creates and maintains the Ray head and worker pods from a single manifest, and Ray Train places one worker per GPU and forms the distributed process group. The training code itself is one function that contains nothing specific to the container, the cluster or the network, so the same script runs unchanged on one node or many.

The fine-tuning target is Qwen2.5-1.5B in bfloat16, which needs about 12 GB for weights, gradients and optimizer state. Sharded eight ways with FSDP, each L4 GPU holds roughly 1.5 GB, and the sample is judged on whether real FSDP collectives crossed both nodes and the job completed. Batches are random token IDs so the run depends on nothing but the cluster; for a real fine-tune, the random tensor is replaced with a Ray Data or DataLoader pipeline and nothing else in the function changes.

Configuration lives in one place, scripts/env.sh, and most changes to the workload touch exactly one file: the model, steps, sequence length and batch size are variables, and an additional Python package is installed at submit time through Ray runtime_env without changing the image. The sample pins train-ml-cuda-v1.1, the current minor release, so a cluster is reproducible from one run to the next, and the scripts are idempotent and safe to re-run.