Reference log-probs for a DPO batch: forward with the adapter disabled (scale 0), materialized + stop_gradient'd so the policy forwards don't hold the reference activations. Returns detached scalars (caller owns).
Reference log-probs for a DPO batch: forward with the adapter disabled (scale 0), materialized + stop_gradient'd so the policy forwards don't hold the reference activations. Returns detached scalars (caller owns).