Version: develop

Writing sflow YAML Configurations

sflow turns one portable recipe — tasks, wiring, resources — into backend-native execution on local / slurm / docker / kubernetes. Write recipes shape-first: get the DAG right, then fill in scripts, scaling, resources, and backend specifics.

Follow the steps below in order. The most common mistake is jumping straight to GPU counts and backend flags before the workflow shape is correct. Don't. Design the DAG first; everything else hangs off it.

Deep field reference → schema-reference.md. Runnable, annotated recipes → examples.md (cited per step). Full docs → nvidia.github.io/nv-sflow/docs/user/intro.

The shared skeleton

Every recipe has the same top-level shape. You design workflow.tasks first; the rest is optional scaffolding you add only when a step calls for it.

version: "0.1"        # schema version (NOT the sflow release) — required in EVERY file
variables: { ... }    # optional — ${{ variables.X }} in YAML, ${X} in scripts
artifacts: [ ... ]    # optional — named paths/URIs, auto-mounted at the same path
backends:  [ ... ]    # optional — omit ⇒ a default `local` backend
operators: [ ... ]    # optional — reusable launch presets
storage:   [ ... ]    # optional — S3 upload targets
workflow:
  name: ...
  tasks: [ ... ]      # THE DAG — design this first (Step 1)

Omit backends/operators and you get local + bash — enough to prototype a DAG on a laptop before touching a cluster.

Step 1 — Shape the DAG (do this first)

The DAG is the recipe. Decide what the tasks are and how they connect before writing any script body.

List the tasks — one node per logical step or service (etcd, frontend, prefill, decode, benchmark, …).
Wire them with depends_on (a task lists the upstreams it waits for). A dependent starts once every upstream is READY (if that upstream has a readiness probe, Step 4) or COMPLETED (if it has none).
Identify the terminal task — the one whose completion ends the run (usually the client/benchmark). Long-running servers never "complete"; sflow stops them when the terminal task finishes.
Reusable fragments wire in reverse with required_by: — a module attaches itself to a shared hub (e.g. benchmark) without editing the hub or adding cross-file depends_on.

Canonical serving shape (Dynamo-style — see examples.md Example 3):

infra (nats, etcd) ──▶ frontend ──▶ workers (prefill + decode | agg) ──▶ benchmark ──▶ end

Write the wiring skeleton first — script bodies come in Step 2:

workflow:
  name: serve_and_bench
  tasks:
    - { name: etcd }
    - { name: frontend,  depends_on: [etcd] }
    - { name: worker,    depends_on: [frontend] }
    - { name: benchmark, depends_on: [worker] }        # terminal task ⇒ run ends here

Sanity-check the shape now: does anything start before its inputs exist? is there exactly one clear end? do "parallel" branches actually need to be concurrent? Fixing shape is cheap here and expensive later.

Step 2 — Write the task scripts

script: is a list of lines joined into one bash script (the python operator instead runs the lines as Python source via python -c).
Don't inline files or hard-code paths. Any file a script needs — launch/helper script, config, model dir, dataset — belongs in an artifact (Step 3), referenced as ${{ artifacts.X.path }}. Never embed Python via heredocs or python3 -c — quoting breaks when YAML → shell → Python nests.
Reference values: ${{ variables.X }} in YAML (plan-time), ${X} in scripts (runtime). For arithmetic/comparison, set the variable's type (integer, …) or it stays a string.

workflow:
  tasks:
    - name: worker
      depends_on: [frontend]
      script:
        - python3 ${{ artifacts.LAUNCH.path }} ${{ variables.MODEL }}   # LAUNCH: see Step 3

See examples.md Examples 1–2.

Step 3 — Handle paths with artifacts (don't hand-manage paths)

An artifact is a named path/URI that sflow resolves, validates, and — inside containers — auto-mounts at the same absolute path on every backend. This is the biggest lever for portability: declare a path once and you stop caring where it lives or how to mount it. ${{ artifacts.X.path }} (YAML, plan-time) and ${X} (script, runtime) always agree, sflow auto-checks the reference at preflight, and auto-mounts it — so you never hand-write container_mounts, docker -v, or k8s volumes for it (doing so is redundant — pitfall 13).

Pick the scheme by what the artifact holds:

Small code snippets & config files → a file:// artifact with inline content:. sflow writes the content and mounts it into the container at the same path on any backend — local, docker, slurm, kubernetes — so you don't manage where it lives or how it gets into the container. Use it for launch/helper scripts, config JSON/YAML, engine specs, etc. Never inline them via heredocs / python3 -c (YAML → shell → Python quoting breaks) or bake them into the image.

artifacts:
  - name: LAUNCH
    uri: file://launch.py             # relative → resolves under SFLOW_WORKSPACE_DIR
    content: |
      import sys; print("serving", sys.argv[1])
  - name: GEN_CONFIG
    uri: file://gen_config.yaml
    content: |
      max_batch_size: 256
      kv_cache_free_gpu_mem_fraction: 0.9
workflow:
  tasks:
    - name: worker
      script:
        - python3 ${{ artifacts.LAUNCH.path }} --config ${{ artifacts.GEN_CONFIG.path }}

Large / pre-existing data (models, datasets, checkpoints) → an fs:// artifact pointing at storage the compute nodes already see. It still resolves + mounts the same path everywhere, but you supply the bytes:
- slurm → keep it on shared storage (Lustre/GPFS/NFS) so the path exists on every node.
- kubernetes → back it with a PVC: declare the PVC under backend volumes: at a mount_path, then point the fs:// artifact at a path under that mount. An fs:// path no PVC covers and the node lacks makes the pod fail (pitfall 9).

Rule of thumb: if a task touches a path, declare it as an artifact — small text → file:// with content:, big data → fs:// on shared storage / a PVC. Artifact schemes & content: rules → schema-reference.md.

Step 4 — Gate the wiring with probes

depends_on alone only waits for a process to start. For services, wait for ready.

Every long-running server gets a readiness probe so dependents wait for real up-ness: log_watch (server logs a ready line), tcp_port (port open), or http_get/http_post (health endpoint).
Add a failure probe on servers — failure.log_watch for "Traceback (most recent call last)" fails the run fast instead of hanging until timeout.
log_watch matches literally by default (parentheses/dots aren't special); prefix re: / regex: for real regex.
Don't probe short-lived one-shot tasks (benchmarks, setup steps).
Defaults: timeout 1200 s (readiness deadline only), interval 5, each_check_timeout 30, success_threshold 1, failure_threshold 3.

- name: frontend
  depends_on: [etcd]
  script: [ "python -m my.frontend --port 8000" ]
  probes:
    readiness: { tcp_port: { port: 8000 } }
    failure:   { log_watch: { regex_pattern: "Traceback (most recent call last)" } }

(On kubernetes, tcp/http probes run from an in-cluster probe pod automatically — the driver host usually can't route to the pod network.) See the probes doc.

Step 5 — Scale with replicas & sweeps

Replicate a task: replicas: { count: N, policy: parallel|sequential }. parallel = all at once (downstream waits for all); sequential = one-by-one (sweeps).
Parameter sweep: replicas.variables: [X] where X declares a domain: → one replica per value (Cartesian product across multiple swept vars). Each replica gets SFLOW_REPLICA_INDEX and the swept values in its env; omit count and it equals the sweep size.

See examples.md Example 5; the replicas doc.

Step 6 — Place on nodes & GPUs

Pin placement with resources.nodes (indices / count / exclude; negative indices count from the end). Request GPUs with resources.gpus.count.
On local/slurm/docker, sflow slices CUDA_VISIBLE_DEVICES per task/replica — set the backend's gpus_per_node so it can pack.
GPU planning: GPUs/worker = TP × DP × PP (framework-dependent — some use attention-DP / expert parallelism), total = GPUs/worker × replicas, nodes = ceil(total / gpus_per_node). When unsure of a framework's parallelism, ask the user.
Use resources.*.release_after (task_ready / task_completion / workflow_completion) to let later tasks reuse a reservation.

See the resources doc.

Step 7 — Backend-specific handling

Pick a backend and apply its rules. Start local, move to a cluster once the DAG is proven.

local — no config needed (default). bash operator. Set nodes: N to simulate a multi-node placement on one machine.

slurm — default operator srun. Backend needs partition, account, time (HH:MM:SS or int minutes), nodes, gpus_per_node. Containers: put container_image / container_name on an srun operator (reuse the image across tasks by container_name). Keep models / workspace / output dir on shared storage (Lustre/GPFS/NFS) so an fs:// artifact auto-mounts at the same path on every node.

backends:
  - name: slurm
    type: slurm
    default: true
    nodes: 1
    partition: ${{ variables.PART }}      # block style: ${{ }} values can't sit in a { } flow map
    account: ${{ variables.ACCT }}
    time: "01:00:00"                       # HH:MM:SS or int minutes
    gpus_per_node: 8
operators:
  - { name: ctr, type: srun, container_image: nvcr.io/nvidia/pytorch:24.07-py3,
      container_writable: true, mpi: pmix }

docker — backend is type: docker, its launch operator is type: docker_run (mind the naming). The image can live on the backend or the operator. Remote/multi-host via a hosts: list.

kubernetes — the workload image lives on the k8s operator, never the backend (the backend reserves + pins nodes with placeholder pods). Rules that bite:

Cluster access is CLI-only (--kubeconfig / --kube-context / --kube-namespace) so recipes stay cluster-agnostic; namespace is a backend field (one namespace per backend).
resources.gpus.count is a per-task total, split evenly across the task's nodes → must be a multiple of the node count. For an ordinary pod sflow leaves CUDA_VISIBLE_DEVICES unset (the device-plugin/DRA assigns the GPUs); when co-located GPU tasks are merged into one pod (merge_colocated_gpu_pods, default auto) it sets a per-member CUDA_VISIBLE_DEVICES to pin each task's slice.
A single-node task → one pod; a resources.nodes task → one pod per node (leader = index 0).
Cluster data → a PVC: declare it under backend volumes: at a mount_path, then point an fs:// artifact at a path under that mount. Output dirs are ephemeral emptyDir — for persistence use a PVC or uploads: (Step 8).
Any MPI workload (mpirun-launched, e.g. TRT-LLM trtllm-llmapi-launch) → type: k8s_mpi, and write the mpirun -np N … explicitly in the script: sflow reads it to set up the MPI world for both single- and multi-node and injects the keypair/hostfile/sshd/-x env forwarding. Reserve plain k8s for non-MPI, single-pod workloads.

backends:
  - { name: k8s, type: kubernetes, default: true, namespace: default, nodes: 1,
      gpus_per_node: 8, scheduling: device_plugin }   # device_plugin default; dra is WIP
operators:
  - { name: gpu_op, type: k8s, image: nvcr.io/nvidia/pytorch:24.12-py3 }

See the backends doc; examples.md Examples 8 (k8s), 10 (k8s + PVC), 11 (slurm + Lustre).

Step 8 — Capture outputs

Metrics: add result: (a regex map, patterns:, or a JSON file:) to parse a task's log/output into ${SFLOW_TASK_OUTPUT_DIR}/result.json plus a workflow results.json. Prefer it over the legacy outputs:. A task isn't COMPLETED until parsing finishes.
Ship files: name storage: S3 targets, then per-task uploads: (fire on COMPLETED) or workflow.upload_all (zip the whole run). Credentials come from the boto3 default chain — never put secrets in YAML; needs the sflow[s3] extra.
Telemetry: monitor: samples GPU/CPU/mem/net on local/slurm/docker (not k8s yet), or enable it without editing the recipe via --enable-workflow-monitor / --enable-task-monitor.

See examples.md Examples 12 (result parsing), 7 (monitor), 9 (uploads); the results / uploads / monitor docs.

Step 9 — Validate before running

Run the one-command preflight before every real run (static schema + reference checks + GPU plan + sflow --dry-run, with a root-cause explanation on failure). Exit 0 = ready.

python scripts/preflight.py my_workflow.yaml
python scripts/preflight.py slurm.yaml common.yaml sglang/agg.yaml --set TP_SIZE=8   # composed

Standalone steps when you want just one:

python scripts/validate_sflow_yaml.py my_workflow.yaml   # schema, refs, k8s volumes
python scripts/check_gpu_plan.py     my_workflow.yaml    # GPU plan + oversubscription
sflow run -f my_workflow.yaml --dry-run                  # full plan + expression resolution

Modular composition (when splitting across files)

Pass fragments directly: sflow run -f slurm.yaml -f common.yaml -f sglang/agg.yaml. Files are combined with a recursive deep merge keyed on name — version must match; variables / artifacts / backends / operators / storage and same-name workflow.tasks are deep-merged (one definition can be scattered across files); leaf scalars/lists are last-file-wins with a warning. Prefer required_by: on optional fragments (absent targets skip silently) over --missable-tasks. See examples.md Example 4.

Pitfalls checklist

Missing version: "0.1" in every file (including fragments).
${{ task.X.nodes }} in a YAML field — task context works only in scripts.
Forgetting depends_on — tasks run immediately without it.
GPU oversubscription: Σ GPU counts > nodes × gpus_per_node.
Probing a short-lived one-shot task (don't probe benchmarks).
Expecting log_watch to be regex — it's literal unless prefixed re: / regex:.
Docker: the operator is type: docker_run (the backend is type: docker).
Putting the workload image on the kubernetes backend — it belongs on the k8s/k8s_mpi operator.
K8s: an fs:// path no volumes: PVC covers and the node lacks → the pod fails (hostPath). Use a PVC.
Expecting K8s task outputs to persist — the output dir is an ephemeral emptyDir; use a PVC or uploads:.
Multi-node Slurm with models/workspace/output on node-local storage — same-path mounts break; use shared storage.
S3 secrets in YAML (use the boto3 chain) or forgetting the sflow[s3] extra.
Hand-writing mounts for a path already declared as an artifact — redundant; artifacts auto-mount same-path.

Additional resources

Field reference & merge rules → schema-reference.md
12 annotated, runnable recipes → examples.md
Full user docs → nvidia.github.io/nv-sflow/docs/user/intro

The shared skeleton​

Step 1 — Shape the DAG (do this first)​

Step 2 — Write the task scripts​

Step 3 — Handle paths with artifacts (don't hand-manage paths)​

Step 4 — Gate the wiring with probes​

Step 5 — Scale with replicas & sweeps​

Step 6 — Place on nodes & GPUs​

Step 7 — Backend-specific handling​

Step 8 — Capture outputs​

Step 9 — Validate before running​

Modular composition (when splitting across files)​

Pitfalls checklist​

Additional resources​