Codeseys/composer-replication-framework

1"""composer_replication.distillation — pluggable self-distillation losses.2 3Per ADR-007, three losses additive to the framework's existing4SDPO/OPSD (`generalized_jsd_loss`):5 6- SimPO: reference-free DPO replacement (channel 3 alternative)7- TAID: annealed teacher interpolation (wraps generalized_jsd_loss for channel 2)8- Entropy-Aware OPD: token-wise gated forward/reverse KL (alternative9  channel-2 wrapper, per ICLR 2026 Spotlight)10 11All three are pure PyTorch — no external deps — so they ship in the core12package without optional extras.13 14Usage in `compose_loss`:15 16    >>> from composer_replication import compose_loss17    >>> components = compose_loss(18    ...     model, batch,19    ...     dpo_variant="simpo",          # channel 3: DPO -> SimPO20    ...     sdpo_wrapper="taid",          # channel 2: SDPO -> TAID21    ...     taid_t=0.4,                   # current TAID interpolation coeff22    ... )23"""24from __future__ import annotations25 26from composer_replication.distillation.simpo import simpo_loss27from composer_replication.distillation.taid import TAIDScheduler, taid_loss28from composer_replication.distillation.entropy_aware_opd import entropy_aware_opd_loss29 30__all__ = [31    "simpo_loss",32    "taid_loss",33    "TAIDScheduler",34    "entropy_aware_opd_loss",35]36