launch-nemo-rl — running NeMo-RL recipes on Kubernetes via nrl-k8s

This is the playbook for the nrl-k8s CLI at infra/nrl_k8s/. Follow it when the user asks to launch / iterate / debug a NeMo-RL recipe on a Kubernetes cluster. Verify current state (kubectl, git log, the recipe + infra files) before acting — the cluster is shared and the cost of a wrong action is high.

1. One command, two modes

There is a single top-level submission command: nrl-k8s run. It has two lifecycle modes.

Ask: Do I need this cluster after the run? If yes, use --raycluster. Otherwise use the default (ephemeral).

The rest of the CLI is observability / stage-by-stage control:

2. Recipe + infra pair

Every launch takes two files. Pass the infra with --infra, not merged inline:

nrl-k8s run infra/nrl_k8s/examples/<recipe>.yaml \
  --infra infra/nrl_k8s/examples/<recipe>.<profile>.infra.yaml

• Recipe (e.g. qwen3_30b_math_8n_4gpu.yaml) — NeMo-RL config: model, GRPO/SFT knobs, cluster.{gpus_per_node,num_nodes}. Uses defaults: to inherit from examples/configs/recipes/llm/....
• Infra (e.g. *.<profile>.infra.yaml) — K8s/Ray shape: namespace, image, service account, RayCluster spec under kuberay:, optional Deployments under deployments:, submit.submitter, launch.{mode,codeSource,codePath,entrypoint}. Pair names follow <recipe>.<profile>[.prod].infra.yaml where <profile> names the hardware target (e.g. gb300).

Example pairs in infra/nrl_k8s/examples/ — read the neighbouring files to see the current conventions for the target profile.

3. Long-lived mode flags

Three independent dimensions. --mode is a macro that picks defaults; individual flags override it.

--mode interactive   → --submitter portForward  --code-source upload  (tails logs)
--mode batch         → --submitter exec         --code-source image   (returns after nohup)

• Submitter: portForward uses kubectl port-forward + Ray Job SDK (gets a submission_id the dashboard tracks). exec uses kubectl exec + nohup on the head pod (no submission_id; driver appears as type=DRIVER in the dashboard).
• Code source: upload stages a working_dir from the laptop (Ray 100 MiB cap). image / lustre expect code on the pod’s filesystem — paired with --code-path (typically /opt/nemo-rl), which is a subPath of the shared-filesystem PVC mount in the standard infra examples.
• Wait: --wait tails logs until terminal; --no-wait returns as soon as the driver is running.

Other long-lived-only flags:

• --replace — stop any running training / daemon job before submitting new ones (suffixes daemon submissionIds with a timestamp so Ray accepts the resubmit).
• --recreate — delete + re-apply a RayCluster whose live spec has drifted from the rendered manifest (default is warn + reuse).
• --skip-daemons — bring up all declared clusters but only submit training. Use on disagg recipes where gym/generation are already healthy.

Gotcha: on infra where the entrypoint does cd /opt/nemo-rl (or another in-image / Lustre path) and loads the recipe from there, --code-source upload does NOT override the recipe on the pod — the uploaded working_dir sits in /tmp/ray/... but the entrypoint cds away from it. To actually test a local recipe change, either sync your edits to the shared filesystem mounted into the pods or flip the Hydra overrides in the entrypoint.

4. Ephemeral mode flags (--rayjob)

When --rayjob is set, run branches into the RayJob code path. Relevant flags:

• --rayjob-name NAME — RayJob metadata name (defaults to the training cluster name).
• --shutdown / --no-shutdown — default true: KubeRay deletes the RayCluster once the Ray Job reaches a terminal state.
• --ttl SECONDS — default 3600s: keep the RayJob object around after the run finishes for post-mortem log access.
• --wait / --no-wait — default wait: poll jobDeploymentStatus until Complete/Failed. --no-wait returns as soon as the RayJob is applied.
• --timeout SECONDS — default 86400s (24h): bound the --wait poll.
• --dry-run — render the RayJob manifest and print it; do not apply.

--replace / --recreate / --skip-daemons are silently ignored in --rayjob mode (KubeRay owns lifecycle).

5. Iterating on a config without touching the shared filesystem

When the recipe on the pod filesystem has the wrong value for your experiment, use Hydra overrides on the entrypoint instead of forking the recipe. Pattern:

entrypoint: |
  set -eu
  cd /opt/nemo-rl
  RUN_ID="\${RAY_JOB_SUBMISSION_ID:-\${NRL_K8S_RUN_ID:-$(date -u +%Y%m%d-%H%M%S)}}"
  python -u examples/run_grpo.py \
    --config infra/nrl_k8s/examples/<recipe>.yaml \
    logger.wandb_enabled=true \
    logger.wandb.project=<project> \
    "logger.wandb.name=<run-name>-\${RUN_ID}"

Escape ${…} with a backslash. OmegaConf otherwise interprets it as interpolation and errors on shell-style ${VAR:-default}. RUN_ID resolves to RAY_JOB_SUBMISSION_ID (injected by KubeRay in rayjob mode) → NRL_K8S_RUN_ID (injected by the CLI in long-lived mode) → local timestamp — so the name is unique across either path.

6. Per-profile concerns (hardware + scheduler + DRA)

Every infra YAML encodes a hardware/scheduler profile. The concrete examples in infra/nrl_k8s/examples/ are authoritative for the profiles they target — read the neighbouring infra file before writing a new one. Things that commonly vary:

• Per-node GPUs (e.g. 4 vs 8) — must match cluster.gpus_per_node in the recipe, otherwise workers stay Pending.
• Node selectors — head pods usually land on a CPU-only node pool; GPU workers match on nvidia.com/gpu.product or a node-group label.
• Scheduler — KAI (schedulerName: kai-scheduler + kai.scheduler/queue label) with topology annotations (kai.scheduler/topology, kai.scheduler/topology-required-placement) gang-schedules workers into one clique. Without it, pods may land on different racks and NVLink/RoCE won’t span them.
• DRA claims — ComputeDomain + RoCE are attached via resourceClaims referencing ResourceClaimTemplates. The CLI auto-creates/deletes these when the worker pod spec contains DRA claim references — no manual setup needed.
• Secrets — always via secretKeyRef (wandb-api-key, image pull secret). Never embed.
• Shared filesystem mounts — typically a Lustre PVC mounted twice: once at the code path (e.g. /opt/nemo-rl with a user-scoped subPath) and once at a workspace root (e.g. /mnt/rl-workspace) for datasets, HF cache, and checkpoints.

Before applying an infra, verify prereqs exist in the target namespace:

kubectl get pvc <workspace-pvc>
kubectl get secret <wandb-secret> <image-pull-secret>
kubectl get sa <service-account>

7. End-to-end workflows

7a. Fresh one-shot run (rayjob)

# From the NeMo-RL repo root:
nrl-k8s check <recipe> --infra <infra>                               # validate first
nrl-k8s run <recipe> --infra <infra> --rayjob --dry-run              # render RayJob manifest
nrl-k8s run <recipe> --infra <infra> --rayjob --no-wait              # apply, returns fast

Watch status + teardown (works even after your laptop disconnects because KubeRay owns the lifecycle):

kubectl get rayjob -n default <name> -w
kubectl get raycluster -n default                                    # empty = teardown succeeded

7b. Dev loop (long-lived)

nrl-k8s run <recipe> --infra <infra> --run-id $(date +%Y%m%d-%H%M%S)
# Edits in the recipe? Just re-run — reuses the live cluster.
# Pod spec changed? Add --recreate to delete + re-apply.
# Disagg recipe with gym/gen already healthy? --skip-daemons.

7c. First-time disaggregated bring-up

nrl-k8s run <recipe> --infra <disagg-infra> --mode batch --code-source image

7d. Cluster-only lifecycle

nrl-k8s cluster up   <recipe> --infra <infra> --target kuberay.training --wait
nrl-k8s cluster up   <recipe> --infra <infra> --target kuberay.training --dry-run   # render manifest
nrl-k8s cluster down <recipe> --infra <infra> --target kuberay.training --wait
nrl-k8s cluster down <recipe> --infra <infra>                                       # tear down all
nrl-k8s cluster list -n default
nrl-k8s cluster dashboard <cluster-name>                                  # port-forward + browser

7e. Deployments (e.g. nemo-skills sandbox)

# Bring up just the deployment
nrl-k8s cluster up <recipe> --infra <infra> --target deployments.nemo_skills
# Tear down just the deployment
nrl-k8s cluster down <recipe> --infra <infra> --target deployments.nemo_skills
# Tear down everything (RayClusters + Deployments)
nrl-k8s cluster down <recipe> --infra <infra>

The deployments: section in infra YAML declares Kubernetes Deployments managed alongside RayClusters. The CLI patches image, imagePullSecrets, and serviceAccountName from the top-level infra keys (same as RayClusters). Deployments start in parallel with cluster bring-up — no ordering dependency.

8. Monitoring a run

# Status
nrl-k8s status <recipe> --infra <infra>
kubectl get rayjob,raycluster -n default

# Follow the driver
nrl-k8s job list <recipe> --infra <infra> --role training
nrl-k8s job logs <run-id> <recipe> --infra <infra> --role training -f

When the nrl-k8s job logs -f subprocess dies (kubectl port-forward i/o timeout after ~15 min idle), just re-run it. The training job keeps going.

To fetch driver logs for a terminal job (SUCCEEDED/FAILED) or a RayJob via the dashboard API:

RC=$(kubectl get rayjob -n default <rayjob-name> -o jsonpath='{.status.rayClusterName}')
kubectl port-forward -n default svc/${RC}-head-svc 18266:8265 &
curl -s http://localhost:18266/api/jobs/                              # lists jobs, find submission_id
curl -s "http://localhost:18266/api/jobs/<submission_id>/logs"        # full driver log

type=DRIVER with submission_id=null means an exec-submitter run (no dashboard log endpoint — use nrl-k8s job logs instead). type=SUBMISSION has submission_id set and /api/jobs/<id>/logs works.

Wandb URL appears in the driver log on the first wandb.init call; grep grep -oE 'https://wandb\.ai/[A-Za-z0-9_./-]+'.

9. Stopping things

Confirm before deleting shared infra. The cost of cluster down on someone else’s cluster is high.

10. Verifying RayJob teardown

After a run --rayjob completes with --shutdown (default), KubeRay should delete the RayCluster:

kubectl get rayjob   -n default <rayjob-name>                        # jobDeploymentStatus = Complete
kubectl get raycluster -n default | grep <rayjob-name>               # no output = torn down

The RayJob object itself sticks around for --ttl seconds (default 3600s) so you can still fetch logs.

11. Common gotchas

• OmegaConf interpolation eats ${VAR} in recipe/infra YAML. Escape shell variables with \${VAR} so OmegaConf passes them through to the pod shell verbatim.
• Megatron optimizer configs don’t carry foreach / fused. Overrides like ~policy.optimizer.kwargs.foreach ~policy.optimizer.kwargs.fused (valid for DTensor configs) break on Megatron recipes. Omit them for Megatron.
• DTensor vs Megatron — MoE recipes typically use megatron_cfg.enabled=true; ensure dtensor_cfg.enabled=false in inherited defaults.
• Shared filesystem vs git divergence — codeSource: image|lustre reads from the pod filesystem. If your local edits aren’t on the shared filesystem the pods mount, the run is testing the on-disk version, not yours. Either sync via a helper pod (head pod exec is often blocked) or override via Hydra flags.
• Ephemeral-storage + readinessProbe are injected by kuberay/CDI webhooks at pod-apply time. Do NOT add them to the inline RayCluster spec.
• Node taints vary per cluster. tolerations: [{operator: Exists}] on workers is defensive and worth keeping.
• Dashboard blank page — Ray 2.52 installs dashboard assets as symlinks by default; nrl-k8s cluster dashboard <name> auto-reinstalls ray[default] --link-mode=copy to fix it. Bake ENV UV_LINK_MODE=copy in the image to avoid this entirely.
• kubectl exec is usually blocked in automation — route around with kubectl get ... -o yaml, kubectl logs, and kubectl port-forward + Ray dashboard APIs.

12. Checklist before calling a run “done”

Before reporting a launch as successful, verify:

1. kubectl get rayjob/raycluster -n default shows the expected objects.
2. nrl-k8s job list (or curl /api/jobs/) shows the job in RUNNING / SUCCEEDED.
3. Driver log contains wandb.ai/<project>/runs/<id> (if wandb is enabled) — share the URL with the user.
4. At least one Processed prompts: 100% line appears (confirms generation is wired).
5. For --rayjob mode only: after jobDeploymentStatus=Complete, confirm kubectl get raycluster | grep <name> is empty (teardown worked).

13. Dev pod

nrl-k8s dev manages a lightweight CPU pod on the cluster for code syncing, debugging, and running kubectl/nrl-k8s from within the cluster.

# One-time: set up secrets (HF token, wandb, SSH key, rclone)
nrl-k8s dev setup-secrets --ssh-key ~/.ssh/id_rsa --add-rclone

# Create pod and exec in (idempotent — reuses existing pod)
nrl-k8s dev connect

# Switch image (must stop first — image change is warned but not auto-applied)
nrl-k8s dev stop
nrl-k8s dev connect --image nvcr.io/nvidian/nemo-rl:v0.7.0

# Tear down
nrl-k8s dev stop

The dev pod:

• Runs on a CPU-only node (anti-affinity to GPU nodes)
• Mounts the shared rl-workspace PVC at /mnt/rl-workspace
• Sets USER env var to the nrl-k8s username (so $USER and getpass.getuser() work correctly despite running as root)
• Installs kubectl, rclone (if configured) on first boot
• Injects SSH keys and tokens via envFrom on a per-user K8s Secret

The pod’s default service account needs an edit RoleBinding in the namespace for kubectl to work inside. dev connect checks this and prints the required YAML if missing.

14. Where things live in the repo

• CLI code: infra/nrl_k8s/src/nrl_k8s/ (cli.py, orchestrate.py, manifest.py, rayjob.py, k8s.py, submitters/, schema.py).
• Tests: infra/nrl_k8s/tests/unit/ — run with uv run --extra test pytest -x -q from infra/nrl_k8s/.
• Recipe + infra examples: infra/nrl_k8s/examples/.
• Base recipes this tool wraps: examples/configs/recipes/llm/… and examples/nemo_gym/….