nvidia-tao/tao-skill-bank/tao-artifacts

tao-artifacts

Four typed artifacts flow through every TAO job. Their schemas live here and nowhere else — producers (model/data skills) and consumers (platform skills) both validate against this skill's references/.

Quick Start — validate an artifact

python - <<'PY'
import json, yaml, jsonschema, pathlib
ref = pathlib.Path("${TAO_SKILL_BANK_PATH:?}/skills/core/tao-artifacts/references")
schema = json.loads((ref / "spec_bundle.schema.json").read_text())
bundle = yaml.safe_load(open("/path/to/bundle.yaml"))   # or a dict built in-context
jsonschema.validate(bundle, schema)                      # raises on violation
print("bundle OK")
PY

Validate the bundle before the verify-before-launch gate; validate a job-record only when debugging (the writer script already enforces the schema).

The two rules the schemas enforce structurally

1. Nested, not dotted. A spec is a nested dict mirroring the container's

config shape — {"train": {"num_epochs": 12}}. Any key containing . at any depth is rejected ({"train.num_epochs": 12} is the #1 authoring mistake). Note the distinction: declared_inputs[].spec_key and gpu_spec_key are dotted/indexed pointers into the spec (dataset.train_data_sources[0].image_dir) — dots are correct there.

1. Mode discrimination. mode: config requires spec + config_format

and a command containing {config_path}, and forbids args. mode: args requires args and forbids spec. There is no other mode.

Optional action lifecycle

Use execution when an action needs more than its primary command. This is the shared model-to-platform seam; do not add a model-specific Docker, Kubernetes, or SLURM renderer merely to carry runtime environment, attestations, post-processing, or helper dependencies.

• The producing model/data skill owns environment, ordered pre_commands,

ordered post_commands, distributed-launch intent, and completion evidence.

• The platform owns container mounts, scheduler/container syntax, task/rank

binding, timeouts, log paths, and preservation of the real child exit code.

• environment is non-secret. Credential values continue to use the selected

platform's secret/sidecar contract and never enter a spec-bundle.

• Commands, environment values, and string values in spec may use

{config_path}, {job_id}, and {results_dir}. The platform binds them only after the job record has been opened; the job record's results_dir is authoritative over any pre-review display path. Persist hashes of both the producer bundle and the bound runtime config.

• supporting_files names checked-in orchestration helpers relative to the

producing skill root. The platform stages the closed set, verifies every declared SHA256, and rejects traversal, undeclared siblings, or overwrite of a different bundle. Supporting files orchestrate an action; they must never shadow or patch code inside the selected image.

• A torchrun declaration expresses process topology, not SLURM/Kubernetes

syntax. Each platform maps it to its native distributed launcher.

Fixed status vocabulary

Every job state anywhere in the pipeline is exactly one of:

PENDING · RUNNING · COMPLETE · ERROR · CANCELED · UNKNOWN

Platform-native sub-states (ImagePullBackOff, PENDING-because-resources, Insufficient-GPU, slurm COMPLETING…) are never new states — they ride in the transition's message field. Terminal = COMPLETE | ERROR | CANCELED. This is what lets the in-turn poll loop and the detached poller share one code path across docker/slurm/kubernetes/brev.

Ordering invariants (enforced at the seam, stated here)

• The verify-before-launch gate runs on the spec-bundle, before any job id

exists.

• tao_job_record.py open writes PENDING + the resolved results_dir

first and returns the id — the only handle a launch can use. A submit that skipped the gate has no id, so it cannot launch.

• transitions is append-only; .tao/ lives outside every synced results

tree.