lens.instruments.rl

View as Markdown

RL metric instruments (rl.* namespace).

Module Contents

Functions

NameDescription
_get_rl_instruments-
record_rl_metricsRecord RL training metrics. All arguments optional; None values skipped.

Data

_RL_INSTRUMENTS

_logger

API

lens.instruments.rl._get_rl_instruments(
meter: opentelemetry.metrics.Meter
) -> dict
lens.instruments.rl.record_rl_metrics(
meter: opentelemetry.metrics.Meter,
reward_mean: float | None = None,
kl_divergence: float | None = None,
policy_loss: float | None = None,
value_loss: float | None = None,
entropy: float | None = None,
response_length_mean: float | None = None,
grad_norm: float | None = None,
learning_rate: float | None = None,
throughput_tokens_per_sec: float | None = None,
generation_duration_ms: float | None = None,
rollout_duration_ms: float | None = None
) -> None

Record RL training metrics. All arguments optional; None values skipped.

lens.instruments.rl._RL_INSTRUMENTS: WeakKeyDictionary = weakref.WeakKeyDictionary()
lens.instruments.rl._logger = logging.getLogger(__name__)