Quick Reference

All sflow.yaml config fields at a glance. The Required column indicates mandatory fields.

For detailed explanations and examples, see Configuration.

Root-Level

Field	Required	Type	Default	Description
version	Yes	string	—	Schema version. Must be "0.1".
variables		dict / list	—	Global variables available to expressions and task env.
artifacts		dict / list	—	Named resources referenced by URI.
backends		dict / list	—	Compute backends (local, slurm, docker, kubernetes).
operators		dict / list	—	Task execution operators (bash, srun, docker_run, ssh, python, k8s, k8s_mpi).
storage		dict / list	—	Post-execution upload targets (S3, …) referenced by task uploads / upload_all (see Storage).
workflow	Yes	object	—	Workflow definition containing name and tasks.

Variables

YAML path: variables.<name>

Field	Required	Type	Default	Description
value	Yes	any	—	Variable value (int, float, bool, string, or list).
description		string	null	Human-readable description.
domain		list	null	Allowed values; enables replica variable sweeps. value must be in domain if set.
type		string	"string"	Type hint (string, integer, etc.).

Artifacts

YAML path: artifacts.<name>

Field	Required	Type	Default	Description
uri	Yes	string	—	Resource URI with scheme (fs://, file://, http://, s3://).
description		string	null	Human-readable description.
content		string	null	Inline file content. Only valid with file:// URI.

Storage

YAML path: storage.<name> — post-execution upload targets referenced by task uploads / upload_all. Credentials come from the provider's default chain (e.g. boto3 for S3) — never inline secrets in YAML. See Uploads.

Field	Required	Type	Default	Description
name	Yes	string	—	Target name (referenced by uploads.target / upload_all.target).
type	Yes	string	—	Target type. Currently s3.

S3 target (type: s3) adds bucket (Yes, string / expr) plus optional region, prefix, endpoint_url (S3-compatible stores such as MinIO / Ceph RGW), storage_class, and addressing_style (auto / virtual / path).

Backends — Common Fields

YAML path: backends.<name>

Field	Required	Type	Default	Description
type	Yes	string	—	local, slurm, docker, or kubernetes.
default		bool	false	Mark as the default backend (only one allowed).
gpus_per_node		int / expr	null	GPUs per node for sflow planning / packing. Does not add Slurm GPU allocation flags.
include_nodes		list[string]	null	Restrict allocation to these node hostnames (unioned with --include-nodes). Translated per backend: Slurm --nodelist, k8s nodeAffinity In, Docker host filter.
exclude_nodes		list[string]	null	Exclude these node hostnames (unioned with --exclude-nodes).

Backends — Local

Additional fields when type: local

Field	Required	Type	Default	Description
nodes		int / expr	1	Number of synthetic local nodes.

Backends — Slurm

Additional fields when type: slurm

Field	Required	Type	Default	Description
account	Yes	string / expr	—	Slurm account.
partition	Yes	string / expr	—	Slurm partition.
time	Yes	string / int / expr	—	Time limit — HH:MM:SS string or integer minutes.
nodes	Yes	int / expr	—	Number of nodes.
gpus_per_node	Yes	int / expr	—	GPUs per node for planning. Set to 0 for CPU-only partitions; tasks that request resources.gpus against a zero-capacity backend will be rejected.
extra_args		list[string]	null	Extra salloc arguments (e.g. --exclusive, --gpus-per-node=8).
job_name		string	null	Job name; defaults to workflow name.

Backends — Kubernetes

Additional fields when type: kubernetes. The workload image lives on the operator (k8s / k8s_mpi), not the backend. Cluster/context/namespace are selected with sflow run flags (--kubeconfig, --kube-context, --kube-namespace), never in YAML. See Backends.

Field	Required	Type	Default	Description
namespace		string / expr	null	Namespace for all pods; must already exist. One namespace per backend. Overridable with --kube-namespace.
nodes		int / expr	1	Number of nodes to reserve.
scheduling		device_plugin|dra	device_plugin	GPU request mode: NVIDIA device plugin (nvidia.com/gpu) or DRA (K8s 1.34+, WIP).
gpu_resource_name		string / expr	nvidia.com/gpu	Device-plugin resource name (e.g. MIG nvidia.com/mig-1g.5gb, or another vendor). Ignored under dra.
gpu_product_label_key		string / expr	nvidia.com/gpu.product	Node label carrying the GPU product string (NVLink-scope auto-detection).
host_network		bool	true	All pods use host networking (pod IP == node IP). Privileged; disable on CNI-routable clusters.
host_ipc		bool	false	Share node IPC namespace + hostPath /dev/shm for cross-pod CUDA IPC over NVLink. Privileged; opt-in.
merge_colocated_gpu_pods		bool | auto|on|disable	auto	Merge co-located single-node GPU tasks into one pod for intra-node NVLink/cuda_ipc.
nvlink_domain		auto|node|rack|disable	auto	Advisory NVLink-domain scope; set only to correct a wrong auto-detection. Not the MNNVL switch (see compute_domain).
rdma		auto|disable|gke|shared_device_plugin|host_device	auto	RDMA/IB provisioning for GPU pods; auto detects the provider, disable forces sockets.
node_selector		map	null	Node-selector labels for all pods. Merge/override with --kube-node-selector.
tolerations		list[map]	tolerate nvidia.com/gpu	Pod tolerations for placeholder + task pods.
image_pull_policy		string / expr	null	Default pod pull policy (Always/IfNotPresent/Never); else cluster default.
probe_pod_image		string / expr	curlimages/curl:latest	Image for the in-cluster TCP/HTTP probe pod (mirror for air-gapped).
cpu_per_gpu		int / expr	null	CPU request per GPU for GPU task pods (requests-only; opt-in).
cpu_request		int / expr	null	CPU request for pods without GPUs (requests-only; opt-in).
collect_max_file_size		int / string	null (10 MiB)	Per-file cap for collecting a task's node-local output back to the driver; 0 disables.
collect_grace_seconds		int / expr	120	Grace period (s) to copy node-local output before pod teardown.
collect_node_local_output		bool	true	Master switch for the node-local output collect. Left on, the DAG waits on a sentinel the pod writes to its log, so a task's output can be lost to log-delivery lag; false runs tasks with no sflow collect machinery in the pod — completion then depends only on pod status, probes and the merge-pod marker, and outputs must reach the driver via a shared filesystem, uploads: or a PVC.
gib_installer_namespace		string / expr	kube-system	Namespace of the GKE gIB (nccl-rdma-installer) DaemonSet.
extra_args		list[string]	null	Extra backend arguments (reserved).
dra		object	null	DRA GPU-allocation options (see below); used when scheduling: dra.
compute_domain		object	null	Multi-Node NVLink (IMEX ComputeDomain) options (see below).
volumes		list[object]	null	PVC / emptyDir volumes mounted into every task pod (see below).
reservation		object	null	Node-reservation tuning (see below).

gpus_per_node (see Common Fields) is derived from node capacity when unset and validated against real GPU capacity at pre-flight; set 0 for CPU-only. PVCs referenced under volumes: must already exist in the namespace — sflow run --kube-skip-pvc drops PVC-backed volumes for debugging.

Kubernetes — volumes[]

YAML path: backends.<name>.volumes[]. Set exactly one of claim or empty_dir.

Field	Required	Type	Default	Description
name	Yes	string	—	Pod volume name (DNS-1123).
mount_path	Yes	string / expr	—	Absolute mount path inside each task pod.
claim	one of	string / expr	null	Existing PVC name. Mutually exclusive with empty_dir.
empty_dir	one of	object	null	Ephemeral per-pod scratch volume. Mutually exclusive with claim.
sub_path		string / expr	null	Path within the volume to mount (volumeMount.subPath).
read_only		bool	PVC true / emptyDir false	Mount read-only.
ensure_writable		bool	false	Inject a root initContainer to chmod the subPath. Requires a claim and read_only: false.

volumes[].empty_dir fields: medium (""|Memory, default "" = node disk; Memory = tmpfs), size_limit (string / expr, e.g. 50Gi).

Kubernetes — dra

YAML path: backends.<name>.dra (used when scheduling: dra; DRA is WIP).

Field	Required	Type	Default	Description
gpu_device_class		string / expr	gpu.nvidia.com	DeviceClass GPUs are requested from.
device_selectors		list[string]	null	CEL expressions narrowing eligible devices (by product/memory).
rdma_device_class		string / expr	null	Opt-in GPU↔NIC co-allocation NIC DeviceClass.
rdma_match_attribute		string / expr	resource.kubernetes.io/pcieRoot	Attribute the GPU and NIC requests must share (PCIe root).
nvlink_domain_label_key		string	null	Node label key identifying the physical NVLink domain (multi-domain placement).

Kubernetes — compute_domain

YAML path: backends.<name>.compute_domain — Multi-Node NVLink (IMEX). Independent of scheduling.

Field	Required	Type	Default	Description
channel		string / expr	null	Join an existing IMEX channel template; auto = the sole existing domain; off/empty = no claim.
create		bool	false	Create a fresh ComputeDomain CR instead of joining. Privileged; needs computedomains RBAC.

Kubernetes — reservation

YAML path: backends.<name>.reservation — tuning for the node reservation.

Field	Required	Type	Default	Description
timeout		int / expr	600	Seconds to wait for every placeholder pod to be scheduled.
handoff		auto|create_before_destroy|destroy_before_create	auto	Placeholder→task handoff order (auto = destroy-before-create when a GPU ResourceQuota is present).
placeholder_image		string / expr	bash:5	Image for reservation placeholder pods (mirror for air-gapped).

Operators — Common Fields

YAML path: operators.<name>

Field	Required	Type	Default	Description
type	Yes	string	—	Operator type: bash, srun, docker_run, ssh, python, k8s, or k8s_mpi.

Operators — srun

Additional fields when type: srun

Field	Required	Type	Default	Description
job_id		string	null	Existing Slurm job ID.
nodes		int / string	null	Node count.
nodelist		list[string]	[]	Node list.
partition		string	null	Slurm partition.
account		string	null	Slurm account.
qos		string	null	QOS.
reservation		string	null	Reservation.
time		string	null	Time limit.
constraint		string	null	Slurm constraint.
exclusive		bool	false	Exclusive node allocation.
chdir		string	null	Working directory.
cpus_per_task		int / string	null	CPUs per task.
gpus		string	null	GPU spec (e.g. all, 1, device=0).
gpus_per_task		string	null	GPUs per task.
gres		string	null	Generic resource spec.
mem		string	null	Memory.
mem_per_cpu		string	null	Memory per CPU.
ntasks		int / string	null	Number of tasks.
ntasks_per_node		int / string	null	Tasks per node.
export		string	"ALL"	Environment export setting.
label		bool	true	Prefix output with task label.
unbuffered		bool	true	Unbuffered output.
kill_on_bad_exit		bool	false	Kill job on non-zero task exit.
overlap		bool	true	Allow step overlap.
wait		int / string	null	Wait time.
container_image		string	null	Container image (Pyxis). Mutually exclusive with container_name.
container_name		string	null	Existing container name (Pyxis). Mutually exclusive with container_image.
container_mount_home		bool	false	Mount home directory in container.
container_writable		bool	true	Writable container filesystem.
container_mounts		list[string]	[]	Bind mounts (e.g. "/host:/ctr:rw").
container_workdir		string	null	Container working directory.
container_remap_root		bool	false	Remap root inside container.
mpi		string	null	MPI type (e.g. pmix, ucx).
extra_args		list[string]	[]	Extra CLI arguments.

Operators — Docker Run

Additional fields when type: docker_run

Field	Required	Type	Default	Description
image	Yes	string	—	Docker image.
workdir		string	null	Working directory inside container.
mounts		list[string]	[]	Bind mounts (e.g. "/host:/ctr:rw").
gpus		string	null	GPU spec (e.g. all, device=0).
extra_args		list[string]	[]	Extra docker run arguments.
pass_envs		bool	true	Forward host environment variables.

Operators — SSH

Additional fields when type: ssh

Field	Required	Type	Default	Description
host	Yes	string	—	SSH host.
user		string	null	SSH user.
port		int	null	SSH port.
identity_file		string	null	Path to identity file.
extra_args		list[string]	[]	Extra SSH arguments.

Operators — Python

Additional fields when type: python

Field	Required	Type	Default	Description
python_exec		string	"python"	Python executable.
extra_args		list[string]	[]	Extra Python arguments.

Operators — Kubernetes (k8s)

Additional fields when type: k8s. The workload image lives on the operator (the kubernetes backend has none). See Backends for the complete pod/container passthrough list.

Field	Required	Type	Default	Description
image	Yes	string	—	Container image for the pod.
image_pull_policy		string	null	Always / IfNotPresent / Never (else cluster default).
restart		string	"Never"	Pod restartPolicy.
host_network		bool	inherit backend	Pod IP == node IP (needed by some RDMA paths).
node_selector		map	inherit backend	Constrain the pod to nodes with these labels.
run_as_root		bool	false	Force runAsUser/Group=0 (root-owned PVCs, MPI/sshd bootstrap).
pass_envs		bool	true	Create + mount the env Secret.
shm_size		string	null	Cap for RAM-backed /dev/shm (e.g. 64Gi); unset = node-RAM-bounded tmpfs.
tty		bool	false	Allocate a pseudo-TTY so \r progress bars stream live.
cpu		int / string	null	Container CPU request.
memory		string	null	Container memory request (e.g. 16Gi).
cpu_limit		int / string	null	Optional CPU hard cap (unset = requests-only).
memory_limit		string	null	Optional memory hard cap (unset = requests-only).
env		list[dict]	null	Raw k8s EnvVar entries (only way to use valueFrom).
image_pull_secrets		list[string]	null	Secret name(s) for pulling from a private registry (e.g. nvcr.io).
service_account		string	null	Pod service account (RBAC / cloud workload identity).
working_dir		string	null	Container working directory.
ports		list[dict]	null	Container ports entries.
tolerations		list[dict]	inherit backend	Pod tolerations (added to the backend's).
collect_max_file_size		int/string	inherit backend	Per-file cap for collecting node-local output back to the driver; 0 disables.
security_context		map	null	Container securityContext (deep-merged; caps union-merged).
device_class / device_selectors		string/list	inherit backend	DRA overrides (ignored under device_plugin).
container_overrides / pod_overrides		map	null	Raw escape hatches, deep-merged last (they win).

Operators — Kubernetes MPI (k8s_mpi)

Additional fields when type: k8s_mpi — any MPI (mpirun) job, single- or multi-node. Inherits every k8s field above and adds an mpi: block. Write an explicit mpirun -np N ... <workload> and sflow builds the MPI world either way, injecting the SSH keypair, hostfile, and launcher wiring. Use plain k8s for non-MPI, single-pod workloads.

mpi. field	Type	Default	Description
route	auto|operator|pods	auto	MPIJob CR (Kubeflow MPI Operator) vs plain-pods bootstrap; auto detects the CRD.
slots_per_worker	int/string	per-node GPU count	Slots advertised per worker.
run_launcher_as_worker	bool	true	Launcher pod also runs a rank.
launcher_creation_policy	AtStartup|WaitForWorkersReady	WaitForWorkersReady	When the launcher starts.
ssh_port	int / expr	2222	Port for the injected sshd.
mpi_implementation	OpenMPI|Intel|MPICH	OpenMPI	Selects hostfile/env conventions.
ensure_sshd	bool	true	Install openssh-server if the image lacks sshd.
forward_env_prefixes	list[string]	[]	Extra env-namespace prefixes forwarded to remote ranks.
omp_num_threads	int	8	Injected OMP_NUM_THREADS per rank; null/0 keeps the image default.
cpu_bind	core|numa|none	core	Per-rank CPU binding, injected only when several ranks share a pod.
cpu_bind_cores_per_rank	int	8	Max cores bound per rank under cpu_bind: core; 0 = uncapped.
worker_setup_timeout_seconds	int / string	900	Per-node worker setup / readiness budget before the launcher starts.
launcher_discovery_timeout	int / string	600	Max wait for the MPI operator to create the launcher pod (route: operator).

Workflow

YAML path: workflow

Field	Required	Type	Default	Description
name	Yes	string	—	Workflow name.
tasks	Yes	list	—	List of task definitions (must be non-empty).
timeout		string / int	null	Not enforced — accepted and merged, but nothing reads it; a workflow that sets it runs unbounded and sflow run warns. Use the backend's own limit (Slurm --time).
variables		dict / list	null	Workflow-scoped variables (same format as root variables).
upload_all		object	null	Zip the whole workflow output dir and upload it to a storage target (see Workflow Upload-All).
monitor		object	null	Workflow-level hardware monitor (see Monitor).

Tasks

YAML path: workflow.tasks[]

Field	Required	Type	Default	Description
name	Yes	string	—	Task name (must be unique).
script	Yes	list[string]	—	Script lines to execute (non-empty).
operator		string / object	null	Operator name, or inline operator override object.
backend		string / dict	null	Backend name, or inline backend override.
depends_on		list[string]	null	Names of tasks this task depends on.
required_by		list[string]	null	Reverse dependency: A required_by: [B] is folded into B depends_on: [A] at load (lets an optional fragment attach to a hub without editing it).
timeout		int / string	null	Not enforced — see Workflow. Bound the task with the backend's own limit instead.
variables		dict / list	null	Task-scoped variables.
resources		object	null	Node / GPU resource requirements.
replicas		object	null	Replication configuration.
retries		object	null	Retry configuration.
probes		object	null	Readiness and failure probes.
outputs		list	null	Output parsing configuration (legacy MVP).
result		map / object	null	Consolidated result parsing (regex map, patterns, or file). Writes result.json + workflow results.json.
fail_fast		bool	backend	Prepend set -e to the script (shell operators) so the task fails on the first failing command. Unset takes the backend default: true on Kubernetes, false on Slurm/local/docker. Set explicitly to override.
uploads		list	null	Per-task file uploads to a storage target, fired on COMPLETED (see Task Uploads).
ports		list	null	Service ports the task exposes, feeding ${{ task.<name>.service }} expressions (see Task Ports).
monitor		object	null	Per-task hardware monitor (see Monitor).

Task Resources

YAML path: workflow.tasks[].resources

Field	Required	Type	Default	Description
nodes.indices		list[int / expr]	null	Specific node indices (e.g. [0]).
nodes.count		int / expr	null	Number of nodes.
nodes.exclude		int / list[int] / expr	null	Node indices to remove from the placement pool before indices, count, or GPU packing.
nodes.release_after		string	null	When node reservations can be reused: workflow_completion, task_ready, or task_completion. sflow reserves nodes only when this is set explicitly; omitted, nodes.indices/count are non-exclusive placement constraints (see note below).
gpus.count	One of count / indices	int / expr	null	Number of GPUs (sets CUDA_VISIBLE_DEVICES). Index-agnostic on its own: the planner packs the task into any contiguous idle GPU run on one node.
gpus.indices	One of count / indices	list[int / expr]	null	Pin the task to specific 0-based, non-negative, unique device ids. Combined with count, count is the total across nodes and indices the per-node slice, so the task fans out over count / len(indices) nodes.
gpus.release_after		string	inferred	When GPU reservations can be reused: workflow_completion, task_ready, or task_completion.

For nodes, release_after only creates an exclusive node reservation when explicitly set; omitted nodes.indices and nodes.count are placement constraints and may overlap with other planned tasks. For GPUs, omitted release_after is inferred: tasks without readiness probes release GPUs after task completion for downstream dependents, while tasks with readiness probes keep GPUs until workflow completion unless explicitly set to task_ready. task_ready releases after readiness succeeds. task_completion releases after any terminal task status (COMPLETED, FAILED, TIMEOUT, or CANCELLED). Dry-run rehearses these resource lifetimes across the DAG.

Task Replicas

YAML path: workflow.tasks[].replicas

Field	Required	Type	Default	Description
count		int / expr	null	Number of replicas.
policy		string / expr	"parallel"	"parallel" or "sequential".
variables		list[string]	null	Variable names for sweeps (Cartesian product of domains).

Task Retries

YAML path: workflow.tasks[].retries

Field	Required	Type	Default	Description
count	Yes	int / expr	—	Number of retries.
interval	Yes	int / expr	—	Delay between retries (seconds).
backoff		int / expr	1	Backoff multiplier.

Task Probes (Readiness / Failure)

YAML path: workflow.tasks[].probes.readiness or workflow.tasks[].probes.failure

Field	Required	Type	Default	Description
delay		int / expr	0	Initial delay before probing (seconds).
timeout		int / expr	1200	Max readiness wait time (seconds). Failure probes do not use this as an overall deadline.
each_check_timeout		int / expr	30	Timeout for a single probe check attempt.
interval		int / expr	5	Check interval (seconds).
success_threshold		int / expr	1	Consecutive successes required.
failure_threshold		int / expr	3	Consecutive failures before failing.

Exactly one probe type must be set per probe:

Probe Type	Required Fields	Optional Fields	Description
tcp_port	port	host, on_node ("first" / "each")	TCP connection check.
http_get	url	headers	HTTP GET health check.
http_post	url	headers, body	HTTP POST health check.
log_watch	regex_pattern or match_pattern	logger, match_count	Match pattern in task logs. Literal by default; prefix with re: or regex: for regular expressions.

Task Outputs

YAML path: workflow.tasks[].outputs[]

Field	Required	Type	Default	Description
pattern	Yes	string	—	Parse pattern (e.g. "TTFT: {ttft:f} ms").
source		string	"stdout"	Log source: stdout or stderr.
metrics.<key>.description		string	null	Metric description.
metrics.<key>.type		string	null	Metric type.
metrics.<key>.aggregate		string	null	Aggregation hint.

Task Result

YAML path: workflow.tasks[].result. Accepts a simple name: regex map, an object with patterns:, or an object with file:. See Results.

patterns and file are mutually exclusive.

Field	Required	Type	Default	Description
patterns		list	null	Advanced regex patterns (see below). Mutually exclusive with file.
file		string	null	Relative path to a JSON source file ending in .json (e.g. result.json). Normalized into result.json.
source		string	"log"	Source selector. Only log is implemented.

result.patterns[]:

Field	Required	Type	Default	Description
name	Yes	string	—	Result key under values.
regex	Yes	string	—	Python regex; prefer one capture group or a named value group.
type		string	"auto"	auto, string, int, float, bool, or json.
unit		string	null	Optional metadata unit (e.g. ms).
aggregate		string	"last"	first, last, list, count, min, max, avg, or sum.
required		bool	false	If true, a missing match marks the result ok: false.
source		string	inherited	Per-pattern source override; only log is implemented.
group		string / int	value then 1	Capture group name or index to extract.

A simple name: regex map is shorthand for patterns with type: auto, aggregate: last, required: false, source: log.

To publish a metric literally named file, use patterns:; a top-level file: key is reserved for JSON file-source results.

Task Uploads

YAML path: workflow.tasks[].uploads[] — copy files to a storage target when the task reaches COMPLETED. See Uploads.

Field	Required	Type	Default	Description
target	Yes	string	—	Name of a storage target.
from	Yes	string	—	Local file or glob to upload.
to		string	null	Remote key/prefix; must end with / when from is a glob.
on_error		warn | fail	warn	What to do if the upload fails.

Workflow Upload-All

YAML path: workflow.upload_all — zip the entire workflow output dir and upload it.

Field	Required	Type	Default	Description
target	Yes	string	—	Name of a storage target.
to		string	${{ workflow.run_id }}.zip	Remote key for the uploaded zip.
on_error		warn | fail	warn	What to do if the upload fails.

Task Ports

YAML path: workflow.tasks[].ports[] — service ports the task exposes, feeding ${{ task.<name>.service }} expressions.

Field	Required	Type	Default	Description
port	Yes	int / expr	—	Port the task serves on.
name		string	null	Optional port name.

Monitor

YAML path: workflow.monitor or workflow.tasks[].monitor — hardware sampling during the run. Monitor fields are concrete ints only (not ${{ }}-resolved). See Monitor.

Field	Required	Type	Default	Description
interval		int	5000	Sampling interval (ms) for built-in scopes without their own.
scopes		object	null (all)	Which scopes to collect — cpu, gpu, memory, disk, network, custom; omit ⇒ all built-ins active.
resources		object	null	Which hardware to target — nodes / gpus (like task resources) or used_by_tasks: [names].
report		object	{enabled: true}	Post-run report. format: any of csv, svg, png (png needs the sflow[monitor] extra).
report.enabled		bool	true	On by default whenever monitor: is set. Set false to opt out — worth doing on large fan-outs, since report cost scales with samples × views. Raw samples and the overview are still written.
window		object	null	{start, end} task-log markers bounding the reported window, so the report covers the benchmark rather than the whole task lifecycle. Task monitors only (rejected on workflow.monitor) and requires report.enabled: true.

Expression Syntax

Fields marked int / expr or string / expr support ${{ ... }} expressions:

Expression	Example
Variable	${{ variables.MY_VAR }}
Variable domain	${{ variables.MY_VAR.domain }}
Backend node IP	${{ backends.slurm_cluster.nodes[0].ip_address }}
Artifact path	${{ artifacts.model_dir.path }}
Task node IP	${{ task.server.nodes[0].ip_address }}

Reserved Environment Variables

Injected by sflow into task environments

These are automatically set by sflow and available in every task script.

Variable	Description
SFLOW_WORKSPACE_DIR	Absolute path to the project workspace root.
SFLOW_OUTPUT_DIR	Global output root directory (default ./sflow_output).
SFLOW_WORKFLOW_OUTPUT_DIR	Output directory for the current workflow run (e.g. sflow_output/<run-id>).
SFLOW_TASK_OUTPUT_DIR	Output directory for the current task replica (e.g. sflow_output/<run-id>/my_task_0).
SFLOW_TASK_RESULT_FILE	Canonical per-task result file path (${SFLOW_TASK_OUTPUT_DIR}/result.json). Write JSON here to publish results directly.
SFLOW_WORKFLOW_RESULT_FILE	Workflow-level results index path (${SFLOW_WORKFLOW_OUTPUT_DIR}/results.json).
SFLOW_REPLICA_INDEX	Zero-based replica index (0, 1, 2, ...).
SFLOW_TASK_ASSIGNED_NODE_NAMES	Comma-separated hostnames of nodes assigned to this task.
SFLOW_TASK_ASSIGNED_NODE_IPS	Comma-separated IP addresses of nodes assigned to this task.
SFLOW_BACKEND_JOB_ID	Backend allocation/job id when available. For Slurm this mirrors SLURM_JOB_ID / SLURM_JOBID.
SFLOW_BACKEND_NODELIST	Backend allocation nodelist when available. For Slurm this mirrors SLURM_JOB_NODELIST / SLURM_NODELIST.
SFLOW_BACKEND_NUM_NODES	Number of nodes in the backend allocation when available. For Slurm this mirrors SLURM_NNODES.
CUDA_VISIBLE_DEVICES	Comma-separated GPU indices allocated to this task (set when resources.gpus.count is used).

In addition, all resolved variables and artifacts paths are injected as environment variables accessible via ${VAR_NAME} in scripts.

Avoid naming a variable after a reserved SFLOW_* / CUDA_VISIBLE_DEVICES env var above — sflow injects and owns these at launch, so a same-named variable collides and causes undefined behavior. sflow run --dry-run prints a Reserved env collisions section listing any such variables so you can rename them before a real run.

Read by sflow from the host environment

sflow reads these to detect an existing Slurm allocation and skip salloc. When a Slurm controller provides SLURM_* / SLURMD_* variables, sflow preserves those controller values in task environments even if workflow variables use the same names.

Variable	Description
SLURM_JOB_ID / SLURM_JOBID	Current Slurm job ID. Used to detect an existing allocation.
SLURM_JOB_NODELIST / SLURM_NODELIST	Node list for the current Slurm allocation.

Provided by Slurm at runtime

These are set by Slurm (not by sflow) and commonly used in task scripts. The srun operator also maps common Slurm step/rank variables into backend-agnostic aliases.

Slurm variable	SFLOW alias	Description
SLURM_STEP_ID	SFLOW_BACKEND_STEP_ID	Slurm step id for the current srun step.
SLURMD_NODENAME	SFLOW_TASK_NODE_NAME	Hostname of the node running the task process.
SLURM_NODEID	SFLOW_TASK_NODE_INDEX	Node index within the allocation (useful for NODE_RANK).
SLURM_PROCID	SFLOW_TASK_PROCESS_ID	Global process/rank id within the step.
SLURM_LOCALID	SFLOW_TASK_LOCAL_PROCESS_ID	Local process/rank id on the node.
SLURM_NTASKS	SFLOW_TASK_NUM_PROCESSES	Number of tasks/processes in the step.
SLURM_SUBMIT_DIR	—	Directory from which the job was submitted.