nemo_gym.artifacts.exporters

Run-metadata exporters.

setup_exporters opens one exporter per configured backend during global config resolution; call sites then fan out through the module-level export_* helpers.

A backend’s module — and with it a multi-hundred-millisecond tracking SDK import — is loaded only once its config says it is wired up. That is why the registry pairs each backend with a config model from config_types rather than asking the exporter class whether it is available.

Submodules

Package Contents

Functions

Name	Description
`_load_exporter_class`	-
`export_metrics`	-
`export_rollouts`	-
`exporter_extras`	The install extras of every exporter `config` fully configures, in registry order.
`get_exporters`	The exporters opened for this process. Empty when no backend is configured.
`load_run_exporter_class`	The post-run exporter class registered as `name`; raises `ConfigError` naming the valid choices.
`load_run_exporter_config_model`	The settings model of the post-run exporter `name`; importing it never loads the exporter’s SDK.
`mark_run_failed`	Record that this process’s evaluation ended without its results, so exporters close the run as failed.
`run_failed`	Whether `mark_run_failed` was called since the exporters were last opened.
`setup_exporters`	Open every backend that is fully configured, and log the run config to each.
`teardown_exporters`	Close all open exporters. Safe to call repeatedly and when none are open.

Data

API

nemo_gym.artifacts.exporters._load_exporter_class(
    class_path: str
) -> type[nemo_gym.artifacts.exporters.base.BaseExporter]

nemo_gym.artifacts.exporters.export_metrics(
    metrics: dict[str, typing.Any],
    step: typing.Optional[int] = None
) -> None

nemo_gym.artifacts.exporters.export_rollouts(
    rollouts: list[dict[str, typing.Any]]
) -> None

nemo_gym.artifacts.exporters.exporter_extras(
    config: typing.Any
) -> tuple[str, ...]

The install extras of every exporter config fully configures, in registry order.

A driver that installs Gym must install these too, or the exporter is skipped at run time.

nemo_gym.artifacts.exporters.get_exporters() -> list[nemo_gym.artifacts.exporters.base.BaseExporter]

The exporters opened for this process. Empty when no backend is configured.

nemo_gym.artifacts.exporters.load_run_exporter_class(
    name: str
) -> type[nemo_gym.artifacts.exporters.base.Exporter]

The post-run exporter class registered as name; raises ConfigError naming the valid choices.

Only gym eval export uses this. A live run builds its exporters in setup_exporters, which treats a missing SDK or a backend that won’t start as a warning and carries on. Here the user asked for this exporter, so a missing SDK is an error that says how to install it.

nemo_gym.artifacts.exporters.load_run_exporter_config_model(
    name: str
) -> type[nemo_gym.config.config_types.RunExportConfig]

The settings model of the post-run exporter name; importing it never loads the exporter’s SDK.

nemo_gym.artifacts.exporters.mark_run_failed() -> None

Record that this process’s evaluation ended without its results, so exporters close the run as failed.

nemo_gym.artifacts.exporters.run_failed() -> bool

Whether mark_run_failed was called since the exporters were last opened.

nemo_gym.artifacts.exporters.setup_exporters(
    global_config_dict: omegaconf.DictConfig
) -> list[nemo_gym.artifacts.exporters.base.BaseExporter]

Open every backend that is fully configured, and log the run config to each.

Replaces any previously opened exporters. A backend that fails to start is skipped with a warning: telemetry must not take the run down with it.

nemo_gym.artifacts.exporters.teardown_exporters() -> None

Close all open exporters. Safe to call repeatedly and when none are open.

nemo_gym.artifacts.exporters.EXPORTER_REGISTRY: dict[str, tuple[type[ExporterConfig], str, str]] = {'wandb': (WANDBConfig, 'nemo_gym.artifacts.exporters.wandb:WandbExporter', 'wan...

nemo_gym.artifacts.exporters._EXPORTERS: list[BaseExporter] = []

nemo_gym.artifacts.exporters._RUN_FAILED = False

nemo_gym.artifacts.exporters.logger = logging.getLogger(__name__)