RayWeightSyncScheme¶

class torchrl.weight_update.RayWeightSyncScheme(strategy: Literal['state_dict', 'tensordict'] = 'state_dict')[source]¶

Weight synchronization for Ray distributed computing.

This scheme uses Ray’s object store and remote calls to synchronize weights across distributed workers (Ray actors).

Each remote collector gets its own transport, following the same pattern as multiprocess collectors.

create_receiver() → WeightReceiver¶

Create a receiver for this scheme.

create_sender() → WeightSender¶

Create a sender for this scheme.

create_transport(pipe_or_context: Any) → TransportBackend[source]¶

Create Ray-based transport for a specific remote collector.

Parameters:: pipe_or_context – The Ray actor handle for the remote collector.
Returns:: RayTransport configured for this specific remote collector.

Docs