sagemaker.train.rft.feedback

sagemaker.train.rft.feedback#

Rollout feedback client for reporting completion and rewards to the RFT Runtime Service.

Classes

RolloutFeedbackClient(metadata)

Client for reporting rollout completion to the RFT Runtime Service.
class sagemaker.train.rft.feedback.RolloutFeedbackClient(metadata: dict[str, Any] | RolloutMetadata)[source]#

Bases: object

Client for reporting rollout completion to the RFT Runtime Service.

Calls the runtime service’s /complete-rollout and /update-reward APIs using bearer token auth.

Example:

feedback = RolloutFeedbackClient(metadata)
feedback.report_complete(reward=0.95)
complete_rollout(status: str = 'ready') → None[source]#

Report trajectory completion to the runtime service.

Parameters:

status – Target status - “ready” for success, “failed” for errors.

report_complete(reward: float | List[float]) → None[source]#

Complete the trajectory and report reward(s).

Convenience method that calls complete_rollout() then update_reward().

Parameters:

reward – The computed reward(s) for this rollout.

report_error(error: str, reward: float | None = None) → None[source]#

Report a rollout error, marking the trajectory as failed.

Parameters:

  • error – Error description.

  • reward – Optional partial reward (defaults to 0.0).

update_reward(reward: float | List[float]) → None[source]#

Report reward(s) to the runtime service.

Parameters:

reward – A single float or list of floats for per-turn rewards.

Contents