torchrl.trainers.algorithms.configs.collectors.MultiAsyncCollectorConfig

class torchrl.trainers.algorithms.configs.collectors.MultiAsyncCollectorConfig(create_env_fn: Any = '???', num_workers: int | None = None, policy: Any = None, policy_factory: Any = None, frames_per_batch: int | None = None, init_random_frames: int | None = 0, total_frames: int = -1, device: str | None = None, storing_device: str | None = None, policy_device: str | None = None, env_device: str | None = None, create_env_kwargs: dict | None = None, collector_class: Any = None, max_frames_per_traj: int | None = None, reset_at_each_iter: bool = False, postproc: ConfigBase | None = None, split_trajs: bool = False, exploration_type: str = 'RANDOM', reset_when_done: bool = True, update_at_each_batch: bool = False, preemptive_threshold: float | None = None, num_threads: int | None = None, num_sub_threads: int = 1, cat_results: Any = None, set_truncated: bool = False, use_buffers: bool = False, replay_buffer: ConfigBase | None = None, extend_buffer: bool = False, trust_policy: bool = True, compile_policy: Any = None, cudagraph_policy: Any = None, no_cuda_sync: bool = False, weight_updater: Any = None, weight_sync_schemes: Any = None, weight_recv_schemes: Any = None, track_policy_version: bool = False, worker_idx: int | None = None, trajs_per_batch: int | None = None, trajs_per_write: int | None = None, init_fn: Any = None, _target_: str = 'torchrl.collectors.MultiAsyncCollector', _partial_: bool = False)

Hydra configuration for MultiAsyncCollector.

MultiAsyncCollector shares its constructor surface with MultiSyncCollector (both forward to the same multi-worker base), so the same kwargs are exposed here.