Architecture

How dockyard_rl fits together at runtime.

• Overview — the three fleets, the async GRPO loop, and the weight-staleness handshake that lets training and generation overlap.

• Cluster, fleets, and placement — bootstrap, fleet specs, placement strategies, STRICT_PACK, and pre-flight validation.

• Weight synchronization — the staleness handshake, the NCCL / IPC / HTTP transports, and the refit seam.

• Experience data plane — moving rollout data between non-colocated fleets without a driver round-trip.