core.trainers.async_grpo.loss
core.trainers.async_grpo.loss
Async GRPO’s per-token losses (loss_type) and their metrics.
compute_loss is TRL’s, with _policy_loss in place of TRL’s clipped objective
and its metrics per loss type.
Classes
| Name | Description |
|---|---|
| AsyncGRPOLossMixin | compute_loss with the loss of args.loss_type. |
AsyncGRPOLossMixin
core.trainers.async_grpo.loss.AsyncGRPOLossMixin()compute_loss with the loss of args.loss_type.
Functions
| Name | Description |
|---|---|
| support_ids | Each position’s top-k/top-p candidate set, padded with -1, which TRL main’s |
support_ids
core.trainers.async_grpo.loss.support_ids(inputs, mask_bool)Each position’s top-k/top-p candidate set, padded with -1, which TRL main’s fused head renormalizes the log-probs over as the sampler did; None without truncation, and on TRL releases, which do not record the sets.