"""Domain-specific circuit template compiler and weight manifest. This emits the matrices/tables used by the registered modules. It is not a general C compiler and does not rewrite pretrained LFM3 FFN neurons. """ import hashlib import torch from .engine import AnalyticEngine,FEATURES def manifest(engine): result={'compiler':'lfmplus-fixed-circuit-templates-1','training_tokens':0, 'spec':engine.spec,'features':{},'stored_tensor_bytes':0} for name in FEATURES: module=engine.layers[name];h=hashlib.sha256();size=0 for key,value in sorted(module.state_dict().items()): raw=value.detach().cpu().contiguous().numpy().tobytes() h.update(key.encode());h.update(raw);size+=len(raw) count=sum(p.numel() for p in module.parameters()) if not count or any(p.requires_grad for p in module.parameters()): raise RuntimeError('Every analytic feature must have frozen registered weights') result['features'][name]={'parameters':count,'tensor_bytes':size,'sha256':h.hexdigest(), 'dtypes':sorted({str(p.dtype) for p in module.parameters()})} result['stored_tensor_bytes']+=size return result def compile_features(spec=None,device='cpu'): engine=AnalyticEngine(spec).to(device).eval() engine.prepare() manifest(engine) return engine