nat.plugins.openpipe.remote_backend

Attributes

logger
_DEFAULT_TIMEOUT

Classes

RemoteBackend

HTTP client that satisfies the art.Backend Protocol against art run.

Functions

_dump_model(→ dict)

Module Contents

logger

_DEFAULT_TIMEOUT

class RemoteBackend(base_url: str)

HTTP client that satisfies the art.Backend Protocol against art run.

base_url

_client

_latest_step: dict[str, int]

_model_inference_name(

model: art.backend.AnyModel,

step: int | None = None,

) → str

async close() → None

async register(model: art.backend.AnyModel) → None

async _get_step(model: art.backend.AnyTrainableModel) → int

async _delete_checkpoint_files(

model: art.backend.AnyTrainableModel,

steps_to_keep: list[int],

) → None

async _prepare_backend_for_training(

model: art.backend.AnyTrainableModel,

config: art.dev.OpenAIServerConfig | None,

) → tuple[str, str]

async _train_model(

model: art.backend.AnyTrainableModel,

trajectory_groups: list[art.trajectories.TrajectoryGroup],

config: art.types.TrainConfig,

dev_config: art.dev.TrainConfig,

verbose: bool = False,

) → collections.abc.AsyncIterator[dict[str, float]]

async train(model: art.backend.AnyTrainableModel, trajectory_groups: collections.abc.Iterable[art.trajectories.TrajectoryGroup], \*\*kwargs: Any) → art.types.TrainResult

abstractmethod _train_sft(

model: art.backend.AnyTrainableModel,

trajectories: collections.abc.Iterable[art.trajectories.Trajectory],

config: art.types.TrainSFTConfig,

dev_config: art.dev.TrainSFTConfig,

verbose: bool = False,

) → collections.abc.AsyncIterator[dict[str, float]]

_dump_model(model: art.backend.AnyModel) → dict