Job environment variables

env | grep -E '^(TUNEPLANE_|NRL_|VOLUMES_DIR|HF_)'

Run that inside a job to see the whole contract. Everything below is set by the control plane; your code reads it and never composes a platform path itself. These are the job-side variables. Server settings are a different list, prefixed the same way — see configuration.

Storage

Variable	What it is
`TUNEPLANE_OUT_DIR`	The one directory whose contents survive the job. Checkpoints, exports, artifacts
`TUNEPLANE_RUN_DIR`	This run’s directory, which `out`, `work` and `logs` sit under
`TUNEPLANE_STORAGE_ROOT`	The deployment’s single configured storage root
`TUNEPLANE_WORK_DIR`	Where the job package was unpacked. Exported by the capsule bootstrap rather than by the spec, because it depends on how the package was delivered
`TUNEPLANE_EXP_DIR`	Your experiment’s directory inside the package
`TUNEPLANE_DATASET_OUT_DIR`	Where to write a dataset the run publishes. Set only with `--output-dataset`
`TUNEPLANE_DATA_CACHE`	The shared dataset cache
`TUNEPLANE_HUB_CACHE`	The cache for hubs with no client library
`HF_HOME`	The Hugging Face cache, inside the storage root

Anything written outside $TUNEPLANE_OUT_DIR disappears with the scratch tree when the container exits.

Topology

Variable	What it is
`TUNEPLANE_CLUSTER_NUM_NODES`	Nodes this job actually has
`TUNEPLANE_CLUSTER_GPUS_PER_NODE`	GPUs per node
`TUNEPLANE_PROFILE_OVERRIDES`	A JSON array of `key=value` entries the framework adapter appends to the training argv
`TUNEPLANE_POOL_TOPOLOGY`	JSON, per-pool: name, series, nodes, gpus_per_node, pin_resource, roles. Absent on a single-pool job
`NRL_PIN_RESOURCE`	The Ray custom resource this job’s series pins to
`CLUSTER_PROFILE`	The profile name the submission used

Quota and the watchdog are measured against TUNEPLANE_CLUSTER_*. Occupying more than that gets the job warned, and can get it stopped. Pass these numbers into torchrun --nproc_per_node or accelerate launch --num_processes.

Reporting

Variable	What it is
`TUNEPLANE_JOB_ENABLED`	`1` when ingest is bound, `0` otherwise
`TUNEPLANE_JOB_ENDPOINT`	Where metrics, logs and samples go. Must be reachable from the GPU node
`TUNEPLANE_JOB_RUN_ID`	This run’s id
`TUNEPLANE_JOB_TOKEN`	A per-run, scoped ingest token
`TUNEPLANE_JOB_JUDGE_ENDPOINT`	The platform judge, for a rubric reward
`TUNEPLANE_JOB_JUDGE_TOKEN`	Its token
`TUNEPLANE_JOB_MONITOR_INTERVAL`	Hardware sampling interval in seconds, default `10`

tuneplane.report reads all of these. With no token it does nothing, which is what you want outside a job. See reporting metrics.

Data and material

Variable	What it is
`VOLUMES_DIR`	The root a Volume is mounted under, read-only. One variable, not one per volume: read `$VOLUMES_DIR/<name>`
`<NAME>_DATA_DIR`	One per referenced dataset, from its name uppercased with `-` and `.` as `_`. So `alice/gsm8k-zh@v2` becomes `GSM8K_ZH_DATA_DIR`

Agent environments

Variable	Set when
`TUNEPLANE_JOB_ENVIRONMENT_DIR`	The protocol is materialised (`nemo-gym`)
`TUNEPLANE_JOB_ENVIRONMENT_URL`	The protocol is served (`openenv`) or remote (`openenv-remote`)
`TUNEPLANE_JOB_ENVIRONMENT_SPLIT`	Always, for a job with an environment. `train` or `eval`, decided by the operation

TUNEPLANE_JOB_ENVIRONMENT_SPLIT has no flag and no override. See agent environments.

Sandbox

Variable	What it is
`TUNEPLANE_JOB_SANDBOX_ENDPOINT`	Where model-generated code runs
`TUNEPLANE_JOB_SANDBOX_TOKEN`	Its token
`TUNEPLANE_JOB_SANDBOX_FUSION_URL`	Emitted only when the bound sandbox speaks Sandbox Fusion
`TUNEPLANE_JOB_SANDBOX_FUSION_CONCURRENCY`	Its concurrency limit

These are optional. Without them, an environment runs tool code in a subprocess inside the container.

Identity and provenance

Variable	What it is
`TUNEPLANE_RECIPE`, `_VERSION`, `_DIGEST`, `_PLUGINS`	Which method this is, exactly
`TUNEPLANE_FRAMEWORK`, `_VERSION`	Which framework build
`TUNEPLANE_RUNTIME_ID`	Which deployment artifact runs it
`TUNEPLANE_CORE_VERSION`	The SDK version that built the spec
`NRL_RUN_ID`, `NRL_SUBMIT_USER`, `RUN_USER`	Run id and who submitted it
`NRL_GIT_COMMIT`, `NRL_GIT_DIRTY`, `NRL_CONFIG_SHA`	Provenance
`NRL_TRAIN_RUN_ID`	On a post-training step, the training run it follows
`NEMO_RL_DIR`	Where NeMo-RL lives in the image. Required for a `nemo-rl` job

The NRL_ prefix is historical and still the real name. Do not rename it in a config.

Secrets

Variable	What it is
`HF_TOKEN`, `HUGGING_FACE_HUB_TOKEN`	Injected when the deployment has one, or when you linked your own account
`CLUSTER_SECRETS_FILE`	A path to a secrets file, for a deployment that injects by path rather than by value

Never bake a token into an image. The platform injects these, and an empty value is dropped rather than set to "" so a ${VAR:-default} in your script still falls back.

Check it worked

echo "$TUNEPLANE_OUT_DIR" "$TUNEPLANE_CLUSTER_GPUS_PER_NODE"
env | grep TUNEPLANE_JOB
ls -la "$VOLUMES_DIR"

A variable that is empty was dropped on purpose, which is not the same as the platform failing to set it. An unknown TUNEPLANE_ or NRL_ variable produces a warning from the cluster-side self-check, because it usually means a typo in a deployment’s extra environment.

JobSpec · Custom training · Server configuration

Get started

Core concepts

Train a model

Evaluate and serve

Agent environments

Extend the platform

Collaborate

Integrations

Reference

Help

Job environment variables

Storage

Topology

Reporting

Data and material

Agent environments

Sandbox

Identity and provenance

Secrets

Check it worked

Next

​Storage

​Topology

​Reporting

​Data and material

​Agent environments

​Sandbox

​Identity and provenance

​Secrets

​Check it worked

​Next

Storage

Topology

Reporting

Data and material

Agent environments

Sandbox

Identity and provenance

Secrets

Check it worked

Next