[ray, doc] fix: make Slurm Ray network interface configurable - #7386

What does this PR do?

This PR makes the network interface used by Ray configurable in the Slurm launch example.

On multi-NIC clusters, users can set RAY_NETWORK_INTERFACE to resolve the IPv4 address of that interface independently on every allocated node. The example then passes the resolved address to ray start --node-ip-address for both the head and worker nodes. If the selected interface does not have an IPv4 address on a node, the script exits with a clear error instead of silently using another interface.

The existing behavior remains unchanged when RAY_NETWORK_INTERFACE is not set. The multinode documentation now explains how to use the option and clarifies that Gloo and NCCL have separate interface settings.

Checklist Before Starting

• Search for similar PRs. Queries: Slurm network interface, RAY_NETWORK_INTERFACE, and node-ip-address with Slurm. No open PR implementing this fix was found.
• Format the PR title as [{modules}] {type}: {description}: [ray, doc] fix: make Slurm Ray network interface configurable.

Test

The following checks were run:

bash -n examples/tutorial/slurm/ray_on_slurm.slurm
git diff --check
python3 tests/special_sanity/check_docs_time_info.py
pre-commit run --all-files --show-diff-on-failure --color=always

The IPv4 resolver was also tested with mocked srun output for both successful CIDR-to-IPv4 extraction and the failure path where the selected interface has no IPv4 address.

The change was validated on a four-node Slurm cluster with a 1 Gbps management interface (lan0) and a 200 Gbps HDR interface (ib0). Before explicitly selecting the interface, Ray actors registered with 10.241.* addresses from lan0. With RAY_NETWORK_INTERFACE=ib0, the Ray head and workers registered with 10.245.* addresses from ib0, and the distributed training job completed successfully.

API and Usage Example

The new setting is optional. Existing users do not need to change their launch command.

On a multi-NIC cluster, select the interface when submitting the job:

RAY_NETWORK_INTERFACE=ib0 \
GLOO_SOCKET_IFNAME=ib0 \
NCCL_SOCKET_IFNAME=ib0 \
sbatch examples/tutorial/slurm/ray_on_slurm.slurm

RAY_NETWORK_INTERFACE controls Ray's node addresses only. GLOO_SOCKET_IFNAME and NCCL_SOCKET_IFNAME remain explicit because they configure separate communication layers.

Design & Code Changes

• Add the optional RAY_NETWORK_INTERFACE environment variable to the Slurm example.
• Resolve the selected interface's IPv4 address on each node through srun on the host.
• Bind both the Ray head and workers with --node-ip-address when the option is set.
• Fail early with the node and interface name when address resolution fails.
• Preserve the original address-discovery behavior when the option is unset.
• Document multi-NIC usage and the boundary between Ray, Gloo, and NCCL settings.

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always.
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow, or explain why this is not feasible: an end-to-end test requires a multi-node Slurm allocation with multiple network interfaces. The shell logic was tested with mocked srun output and validated on a real four-node Slurm cluster.
• Once the PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace or the Feishu group.
• Recipe submodule update: not applicable.

AI assistance disclosure

OpenAI Codex assisted with the implementation and documentation. I reviewed every changed line and ran the validation described above.