tensorplay.distributed.launcher API

Functions 1

launch_agent

tensorplay.distributed.launcher.launch_agent(config: LaunchConfig, entrypoint: Callable | str | None, args: list[Any]) → Any[source]

Start the local agent for entrypoint and run it to completion.

Raises ChildFailedError when any worker fails after exhausting the restart budget.

Classes 2

elastic_launch

classFull reference ↗

class tensorplay.distributed.launcher.elastic_launch(config: LaunchConfig, entrypoint: Callable | str | None)[source]

Callable wrapper around launch_agent().

Usage:

elastic_launch(config=LaunchConfig(...), entrypoint="train.py")("--epoch", "10")

LaunchConfig

classFull reference ↗

class tensorplay.distributed.launcher.LaunchConfig(min_nodes: int = 1, max_nodes: int = 1, nproc_per_node: int = 1, run_id: str = '', role: str = 'default_role', rdzv_endpoint: str = '', rdzv_backend: str = 'static', rdzv_configs: dict[str, ~typing.Any]=<factory>, rdzv_timeout: int = -1, max_restarts: int = 0, monitor_interval: float = 0.1, start_method: str = 'spawn', log_dir: str | None = None, redirects: Any = None, tee: Any = None, metrics_cfg: dict[str, str]=<factory>, local_addr: str | None = None, node_rank: int = 0, master_addr: str | None = None, master_port: int | None = None)[source]

Elastic launch settings for one job.

min_nodes/max_nodes bound the job size; when they differ the job is elastic and nodes may join or leave within those bounds. rdzv_* fields select and configure the rendezvous backend.

Functions 1

launch_agent

Classes 2

elastic_launch

LaunchConfig

Help improve this page

On this page