"""End-to-end inference & Cognitive Fabric verification demo for Supermix v91 (Male-CNS Connectome).""" from __future__ import annotations import argparse import sys from pathlib import Path ROOT = Path(__file__).resolve().parent SOURCE_DIR = ROOT / "source" if str(SOURCE_DIR) not in sys.path: sys.path.insert(0, str(SOURCE_DIR)) import torch # noqa: E402 import mimomix_decoding as decoding # noqa: E402 import mimomix_text as text_utils # noqa: E402 from mimomix_core import MiMoMixConfig, MiMoMixModel # noqa: E402 from scf import CognitiveFabricRuntime, SCFFeatureFlags # noqa: E402 from scf.verification import SCFVerificationGate # noqa: E402 DEFAULT_CHECKPOINT = ROOT / "weights" / "v91_cns_connectome.selected.pt" def load_checkpoint(path: Path = DEFAULT_CHECKPOINT): payload = torch.load(path, map_location="cpu", weights_only=False) model = MiMoMixModel(MiMoMixConfig(**payload["config"])) model.load_state_dict(payload["state_dict"]) model.eval() tokenizer = text_utils.WordTokenizer.from_dict(payload["tokenizer"]) return model, tokenizer @torch.no_grad() def generate_reply(model: MiMoMixModel, tokenizer, message: str, max_new_tokens: int = 64) -> str: ids, _ = tokenizer.encode_turn(message, None) result = decoding.speculative_generate( model, torch.tensor([ids]), max_new_tokens=max_new_tokens, eos_token_id=text_utils.EOS, ) return tokenizer.decode(result.new_tokens[0].tolist()).strip() def main() -> int: parser = argparse.ArgumentParser(description="Run Supermix v91 (Male-CNS Connectome) demo") parser.add_argument("--checkpoint", type=Path, default=DEFAULT_CHECKPOINT) args = parser.parse_args() model, tokenizer = load_checkpoint(args.checkpoint) params = sum(p.numel() for p in model.parameters()) print(f"Loaded {args.checkpoint.name}: {params:,} params | vocab {tokenizer.vocab_size}") prompts = [ "nums = [10, 12, 4]\nr = nums[1] + nums[2] Final r?", "How many Li20 neurons does the male CNS contain?", "inverse of [[4, 7], [2, 6]]", "eigenvalues of [[4, 1], [2, 3]]", ] gate = SCFVerificationGate() runtime = CognitiveFabricRuntime() flags = SCFFeatureFlags.all_enabled() for q in prompts: neural = generate_reply(model, tokenizer, q) ver = gate.verify(query=q, candidate_answer=neural) rt = runtime.run_task(q, flags_override=flags) tele = model.telemetry() print(f"\nQ: {q}") print(f"Neural: {neural}") print( f"SCF verified={ver['verified']} | answer={ver['verified_answer']} " f"| source={ver['verifier_source']} | family={rt['routing']['selected_family']}" ) print( f"Telemetry: mhc={bool(tele.get('mhc'))} | engram={bool(tele.get('engram'))} " f"| deltanet={bool(tele.get('deltanet'))} | cns={bool(tele.get('cns_core'))}" ) return 0 if __name__ == "__main__": raise SystemExit(main())