"""Pinned OpenTau core in the shared OpenPI expert/data/training pipeline. This is a common-protocol baseline, not an execution of OpenTau's full trainer. The upstream core in reference_ttt.py is byte-identical to revision 64544b9. """ from dataclasses import replace import torch from torch import nn from .core import FastState,TTTConfig from .reference_core_adapter import ReferenceLayer from .multilayer_adapter import MultilayerExpert class ReferenceExpert(MultilayerExpert): def __init__(self,model,indices,register_config,**kwargs): super().__init__(model,indices,register_config,**kwargs) device=next(model.action_in_proj.parameters()).device self.ttt=nn.ModuleDict({str(i):ReferenceLayer().to(device) for i in indices}) def insert(self,idx,x,incoming,candidate,enabled,differentiable,trace,taps): key=str(idx) if not enabled or key not in self.ttt:return x if idx in self._calls:raise RuntimeError('Duplicate TTT reference call') self._calls.append(idx) if self.observer:self.observer('inner','attempt') if self.layer_observer:self.layer_observer(key,'attempt') result,candidate[key]=self.ttt[key].step(x,incoming[key],differentiable=differentiable) if self.gate_zero or idx in self.zero_read_layers:result=x+0.*(result-x) if self.guard:self.guard({key:candidate[key]}) if not bool(torch.isfinite(result).all()):raise FloatingPointError('Nonfinite reference output') if trace is not None:trace.append(dict(layer=idx,method='opentau_reference',updates=candidate[key].updates.tolist())) if self.layer_observer:self.layer_observer(key,'complete') if self.observer:self.observer('inner','complete') return result