nemo_gym.orchestration.export_step

One post-run export of a submitted benchmark, with its outcome kept in a result file.

The driver step of a job runs this after gym eval run, once per configured export: python3 -m nemo_gym.orchestration.export_step <exporter> +key=value .... The exporters themselves know nothing about it; what a job needs on top of them (telling several exports apart afterwards, not repeating one that finished, retrying one that was interrupted) lives here.

Module Contents

Functions

Name	Description
`_read_state`	-
`_result_payload`	-
`_write_json_atomically`	-
`main`	`python3 -m nemo_gym.orchestration.export_step <exporter> +key=value ...`: the settings are Hydra overrides.
`run_export`	Run the post-run exporter `name` on the run `rollouts_jsonl_fpath` points at.

Data

RUNNING_STATE

_FINAL_STATES

API

nemo_gym.orchestration.export_step._read_state(
    path: pathlib.Path
) -> typing.Optional[str]

nemo_gym.orchestration.export_step._result_payload(
    result: nemo_gym.artifacts.exporters.base.ExportResult
) -> dict[str, typing.Any]

nemo_gym.orchestration.export_step._write_json_atomically(
    path: pathlib.Path,
    payload: dict[str, typing.Any]
) -> None

nemo_gym.orchestration.export_step.main() -> None

python3 -m nemo_gym.orchestration.export_step <exporter> +key=value ...: the settings are Hydra overrides.

nemo_gym.orchestration.export_step.run_export(
    name: str,
    global_config_dict: omegaconf.DictConfig
) -> nemo_gym.artifacts.exporters.base.ExportResult

Run the post-run exporter name on the run rollouts_jsonl_fpath points at.

result_json_fpath is a setting of this step, not of the exporters. When it is set, the outcome is also kept in that file, which is how a job that runs several exports can tell them apart afterwards: no file means never attempted, running means interrupted (for example at the walltime), and anything else is a final result. An export that already succeeded or was skipped is not repeated; a failed or interrupted one is retried. A configuration error is recorded as a failed result too, and still raised.

nemo_gym.orchestration.export_step.RUNNING_STATE = 'running'

nemo_gym.orchestration.export_step._FINAL_STATES = (ExportState.SUCCEEDED.value, ExportState.SKIPPED.value)