fix(configs): K8S required_resources.memory must be bytes, not "8Gi"

[elliot-barn]

Stacked on #940 — targets add-azure-declarative-compute-config so the diff is just this fix.

Problem

The three K8S configs specify memory as Kubernetes quantity strings (8Gi, 16Gi, 32Gi). That form is accepted by anyscale compute-config create -f, because the SDK converts suffixes to bytes before building the API request (PhysicalResources.to_dict(for_api=True) → _parse_memory_string). That's the path #940 was validated on, so it looked fine.

The publish path never touches the SDK. rayapp build copies required_resources verbatim into the bundle, and the console clone path (workspaces_service._make_template_compute_config) parses that bundle straight into the backend's PhysicalResources, whose memory is an Optional[int] of bytes with no quantity parsing.

Confirmed against the live Azure staging API by replaying the built bundle through the legacy create path:

422 Unprocessable Content from POST /cluster_computes/
  body.config.head_node_type.required_resources.memory: value is not a valid integer
  body.config.worker_node_types.0.required_resources.memory: value is not a valid integer

This is latent today only because the launch path force-selects AWS on Kubernetes clouds — it would surface the moment the K8S key is honored, i.e. exactly when #940 becomes useful.

Fix

Integer bytes in all three configs (8Gi = 8589934592, 16Gi = 17179869184, 32Gi = 34359738368). Accepted by both paths — the SDK takes Union[str, int].

Plus a validator rule, because nothing else in the pipeline catches this: the one path that converts suffixes is the path templates are tested on, so a string passes local testing and fails only at launch.

Testing

• pre-commit run --all-files green; validator reports OK: validated 56 BUILD.yaml entries.
• The new rule fires on a reverted value: configs/getting-started/k8s.yaml: worker_nodes[0].required_resources.memory must be an integer number of bytes, not '32Gi'.
• All three bundles rebuilt; each carries integer memory.
• The rebuilt pytorch-fsdp bundle — GPU plus required_labels — now creates successfully through the legacy path on Azure staging (cpt_6ji34wrn5c2iqf43v6h68rdvw2), so the backend's check_gpu_accelerator_consistency passes too.
• End to end: a workspace launched on the byte-valued config on an AKS cloud (compute_stack: K8S, westus2) and workspace-intro ran its full notebook test there — 24 cells, exit 0.

One caveat on that last line: rayapp test can't yet drive this itself on Kubernetes clouds. It reports success while executing nothing (ray-project/rayci#496 fixes the detection), and two further k8s gaps — run_command starting in $HOME rather than the workspace working directory, and bash -c not sourcing the ~/.bashrc where k8s workspaces declare PATH and ANYSCALE_* — had to be worked around by hand to get the run above. None of them are caused by these configs.

🤖 Generated with Claude Code

[Aydin-ab]

This looks like the same issue we had for compute configs few months ago, how did we solve it then ?

Can we maybe install the sdk and use it to convert the config inside of rayapp ?

The goal is to have these .yaml be the actual compute config we document in our docs
https://docs.anyscale.com/configuration/compute/declarative#examples