nemo_rl.models.generation.rollout_result

One rollout’s result and its JSON wire form, for every generation backend.

An engine served through Dynamo sends the result to the driver in engine_data; the vLLM worker builds it with request_pipeline.rollout_result.

Module Contents

Classes

RolloutResult

One rollout’s outcome, independent of where its output is assembled.

Functions

rollout_result_to_wire	JSON-safe form of result, for engines that answer over HTTP.
rollout_result_from_wire	Inverse of :func:rollout_result_to_wire; routes keep their own dtype.

API

class nemo_rl.models.generation.rollout_result.RolloutResult

One rollout’s outcome, independent of where its output is assembled.

token_ids: list[int]

None

logprobs: list[float]

None

truncated: bool

None

routed_experts: Optional[torch.Tensor]

None

r3_stats: Optional[dict[str, int]]

None

classmethod nothing_generated() → nemo_rl.models.generation.rollout_result.RolloutResult

The result when no new token fits: the prompt alone, untruncated.

nemo_rl.models.generation.rollout_result.rollout_result_to_wire(

result: nemo_rl.models.generation.rollout_result.RolloutResult,

) → dict[str, Any]

JSON-safe form of result, for engines that answer over HTTP.

nemo_rl.models.generation.rollout_result.rollout_result_from_wire(

wire: dict[str, Any],

) → nemo_rl.models.generation.rollout_result.RolloutResult

Inverse of :func:rollout_result_to_wire; routes keep their own dtype.