`nemo_rl.experience.rollout_reassembler_actor`#

CPU Ray actors for metadata-only token-capture finalization.

Module Contents#

Classes#

`ReassemblyRequest`	Metadata-only input for one prompt group’s finalization.
`RolloutReassemblerActorConfig`	Internal constructor values shared by every finalizer actor.
`RolloutReassemblerActor`	Own a connect-only TQ client and lightweight finalizer in one process.

Functions#

create_rollout_reassembler_actors

Construct and validate the pool after TQ partitions are registered.

API#

class nemo_rl.experience.rollout_reassembler_actor.ReassemblyRequest#

Metadata-only input for one prompt group’s finalization.

group_id: str#: None

rollout_ids: tuple[str, ...]#: None

canonical_sample_ids: tuple[str, ...]#: None

receipts: tuple[Optional[dict[str, Any]], ...]#: None

rewards: tuple[float, ...]#: None

fallback_weight_version: int#: None

prompt_idx: int#: None

mask_sample: tuple[bool, ...]#: None

loss_multiplier: float#: 1.0

logical_selections: Optional[tuple[nemo_rl.experience.rollout_reassembler.RolloutSelection, ...]]#: None

execution_row_multiple: int#: 1

reward_checks: tuple[nemo_rl.experience.reward_penalties.RewardChecks | None, ...] | None#: None

reward_log_contexts: tuple[nemo_rl.experience.reward_penalties.RewardLogContext | None, ...] | None#: None

resolved_agent_name: Optional[str]#: None

attribution_key: Optional[str]#: None

property capture_receipts: tuple[Optional[dict[str, Any]], ...]#: Return cleanup receipts only after validating Forest capture ownership.

property cleanup_sample_ids: tuple[str, ...]#: Return the segment publication plan, including exact padding.

class nemo_rl.experience.rollout_reassembler_actor.RolloutReassemblerActorConfig#

Internal constructor values shared by every finalizer actor.

partition_id: str#: None

staging_partition: str#: None

pad_token_id: int#: None

router_replay_enabled: bool#: None

defer_routed_experts_to_policy: bool#: None

max_seq_len: int#: None

capture_media: bool#: None

media_columns: Optional[nemo_rl.data.captured_media.MediaColumnSpec]#: None

top_logprobs: int#: None

reward_penalty_config: nemo_rl.algorithms.grpo.RewardPenaltyConfig | None#: None

effort_config: nemo_rl.experience.rollouts.EffortLevelsConfig | None#: None

tokenizer_config: nemo_rl.models.policy.TokenizerConfig | None#: None

thinking_tags: tuple[str, ...] | None#: None

class nemo_rl.experience.rollout_reassembler_actor.RolloutReassemblerActor( dp_config: nemo_rl.data_plane.DataPlaneConfig, config: nemo_rl.experience.rollout_reassembler_actor.RolloutReassemblerActorConfig, )#

Own a connect-only TQ client and lightweight finalizer in one process.

Initialization

mooncake_checkpoint( body: dict[str, Any], ) → dict[str, Any] | None#: Run an owner-local checkpoint command; return metadata, never payloads.

get_data_plane_snapshot() → dict[str, Any] | None#

check_dependencies() → None#: Import the finalization API before the controller starts rollouts.

finalize( request: nemo_rl.experience.rollout_reassembler_actor.ReassemblyRequest, ) → nemo_rl.experience.rollout_reassembler.FinalizedGroup#: Finalize one request without allowing tensor payloads across Ray RPC.

nemo_rl.experience.rollout_reassembler_actor.create_rollout_reassembler_actors( dp_config: nemo_rl.data_plane.DataPlaneConfig, config: nemo_rl.experience.rollout_reassembler_actor.RolloutReassemblerActorConfig, *, num_workers: int, ) → list[Any]#: Construct and validate the pool after TQ partitions are registered.

nemo_rl.experience.rollout_reassembler_actor#

Module Contents#

Classes#

Functions#

API#

`nemo_rl.experience.rollout_reassembler_actor`#