"""Isolated donor ALIGN-NEW; no change to the ALIGN v1/GRASS adapter.""" import math import torch from . import register_algorithm from .mp_opd import MPOnPolicyKD @register_algorithm('mp_opd_align_new') class AlignNewMPOnPolicyKD(MPOnPolicyKD): # Let the unchanged parent initialize the same chunk aligner as ALIGN v1. mode = 'align' def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self.mode = 'align_new' self.rho = float(self.kd.mp_opd_align_new_min_descent_fraction) self.eps_g = float(self.kd.mp_opd_align_new_eps_g) if self.kd.mp_opd_grass_geometry != 'exact_head': raise ValueError('ALIGN-NEW requires exact_head geometry') if not math.isfinite(self.rho) or not 0 <= self.rho <= 1: raise ValueError('ALIGN-NEW minimum descent fraction must be in [0,1]') if not math.isfinite(self.eps_g) or self.eps_g < 0: raise ValueError('ALIGN-NEW eps_g must be finite and nonnegative') def _effective_rates(self, credits, atoms, partition, hidden, labels, **_kwargs): from ._mp_opd_port._mp_opd_align import align_chunks, align_metrics from ._mp_opd_port._mp_opd_align_new import align_new_guard from ._mp_opd_port._mp_opd_grass_chunk import chunk_head_gram from ._mp_opd_port._mp_opd_head_pair import head_pair_geometry ranges = tuple((a.student_start, a.student_end) for a in atoms) with torch.no_grad(): logits = self._head_logits(self.student.model.lm_head, hidden, self.student_cap, True) has_bias = getattr(self.student.model.lm_head, 'bias', None) is not None # Reconstruct the same logits as the credit path before disabling # ambient autocast for donor geometry/local coefficient arithmetic. with torch.autocast(device_type=hidden.device.type, enabled=False): head = chunk_head_gram(logits, hidden.detach(), labels, ranges, partition, selected_log_prob=-credits.student_token_nll, softcap=self.student_cap, head_bias=has_bias, diagonal_only=False) stats = align_chunks(credits.rate, head.gram, partition, eps_h=self.kd.mp_opd_align_eps_h) candidate = stats['rate'].to(credits.rate.dtype) metrics = align_metrics(credits.rate, candidate, head.gram, stats) pair = head_pair_geometry(logits, hidden.detach(), labels, ranges, credits.rate, candidate, selected_log_prob=-credits.student_token_nll, softcap=self.student_cap, head_bias=has_bias) effective, global_metrics = align_new_guard(credits.rate, candidate, pair.first_norm2, pair.second_norm2, pair.dot, min_descent_fraction=self.rho, eps_g=self.eps_g) metrics = {k.replace('mp_opd_align_', 'mp_opd_align_new_local_'): v for k, v in metrics.items()} metrics.update(global_metrics) return effective, metrics