"""Use pinned OpenPI conversion with a strict, recorded key audit. No missing inference weight is accepted. The expert's unused vocabulary head is zero-filled explicitly; tied aliases are populated from their source weight. """ import argparse,hashlib,json,pathlib,sys,subprocess,dataclasses def main(): p=argparse.ArgumentParser();p.add_argument('--openpi-source',required=True);p.add_argument('--checkpoint-dir',required=True);p.add_argument('--output',required=True) p.add_argument('--model-profile',choices=('pi05_base_libero_protocol','legacy_pi05_libero'),default='pi05_base_libero_protocol') args=p.parse_args();source=pathlib.Path(args.openpi_source).resolve() revision=subprocess.check_output(['git','-C',str(source),'rev-parse','HEAD'],text=True).strip() if revision!='215abfb217dbac7d5f1273282331b9b1866c0479':raise ValueError('Unvalidated OpenPI revision') subprocess.run(['git','-C',str(source),'diff','--quiet'],check=True) sys.path[:0]=[str(source/'src'),str(source/'packages/openpi-client/src')] import torch from openpi.models.pi0_config import Pi0Config script=source/'examples/convert_jax_model_to_pytorch.py' text=script.read_text().replace('import openpi.training.config as _config','') needle=' pi0_model.load_state_dict(all_params, strict=False)' if text.count(needle)!=1:raise ValueError('Pinned converter changed') text=text.replace(needle,' audit_and_load(pi0_model, all_params)') save_needle=' safetensors.torch.save_model(pi0_model,' if text.count(save_needle)!=1:raise ValueError('Pinned converter save changed') text=text.replace(save_needle,' apply_runtime_precision(pi0_model)\n'+save_needle) out=pathlib.Path(args.output);out.mkdir(parents=True,exist_ok=True) if (out/'model.safetensors').exists():raise FileExistsError('Converted output already exists') checkpoint_name='pi05_libero' if args.model_profile=='legacy_pi05_libero' else 'pi05_base' audit={'source_revision':revision,'saved_dtype':'float32_before_runtime_mixed_cast','source_url':'gs://openpi-assets/checkpoints/'+checkpoint_name,'model_profile':args.model_profile,'converter_sha256':hashlib.sha256(script.read_bytes()).hexdigest(),'tied_aliases':{},'unused_zero_filled':[]} def load(model,state): # Preserve original FP32 tensors before the runtime mixed-precision cast. model.to(torch.float32) expected=model.state_dict();groups={} for n,v in expected.items():groups.setdefault(v.data_ptr(),[]).append(n) for names in groups.values(): supplied=[n for n in names if n in state] if supplied: for name in names: if name not in state: state[name]=state[supplied[0]];audit['tied_aliases'][name]=supplied[0] missing=set(expected)-set(state);unexpected=set(state)-set(expected) unused={'paligemma_with_expert.gemma_expert.lm_head.weight'} audit.update(missing_before_unused_fill=sorted(missing),unexpected=sorted(unexpected)) (out/'loading-audit.json').write_text(json.dumps(audit,indent=2)) if unexpected or missing-unused:raise ValueError('Conversion has missing/unexpected inference weights: '+str((missing,unexpected))) for n in missing: state[n]=torch.zeros_like(expected[n]);audit['unused_zero_filled'].append(n) model.load_state_dict(state,strict=True) audit.update(strict_load=True,tensor_count=len(state),finite=all(bool(torch.isfinite(v).all()) for v in state.values())) if not audit['finite']:raise ValueError('Nonfinite converted weights') (out/'loading-audit.json').write_text(json.dumps(audit,indent=2)) def runtime_precision(model): if args.model_profile=='legacy_pi05_libero': model.paligemma_with_expert.to_bfloat16_for_selected_params('bfloat16') audit['saved_dtype']='runtime_mixed_bfloat16_float32' (out/'loading-audit.json').write_text(json.dumps(audit,indent=2)) ns={'__name__':'pinned_openpi_conversion','audit_and_load':load,'apply_runtime_precision':runtime_precision} exec(compile(text,str(script),'exec'),ns) legacy=args.model_profile=='legacy_pi05_libero' config=Pi0Config(pi05=True,action_horizon=10,max_token_len=200 if legacy else 256,discrete_state_input=not legacy,pytorch_compile_mode=None) audit['architecture']=dataclasses.asdict(config) ns['convert_pi0_checkpoint'](str(pathlib.Path(args.checkpoint_dir).resolve()),'float32',str(out),config) if __name__=='__main__':main()