The existing DPO/ORPO head geometry, shared with inference. Two- and three-row response spans pad to four before the head and slice back. Borrows hidden and model, returns caller-owned logits; preserves training numerics.
The existing DPO/ORPO head geometry, shared with inference. Two- and three-row response spans pad to four before the head and slice back. Borrows hidden and model, returns caller-owned logits; preserves training numerics.