| """Export complete review artifacts to lossless non-executable tensor trees.""" |
| from pathlib import Path |
| import argparse,hashlib,json,shutil,time |
| import torch |
| from safe_tree import save_tree,load_tree |
|
|
|
|
| def sha(path):return hashlib.sha256(path.read_bytes()).hexdigest() |
|
|
|
|
| def compare(a,b): |
| if torch.is_tensor(a):assert torch.is_tensor(b) and torch.equal(a,b);return 1 |
| if isinstance(a,dict):assert list(a)==list(b);return sum(compare(a[k],b[k]) for k in a) |
| if isinstance(a,(tuple,list)):assert type(a)==type(b) and len(a)==len(b);return sum(compare(x,y) for x,y in zip(a,b)) |
| assert a==b;return 0 |
|
|
|
|
| def main(): |
| parser=argparse.ArgumentParser();parser.add_argument('--source',default=str(Path(__file__).resolve().parent));parser.add_argument('--output',required=True);parser.add_argument('--device',default='cuda:3');args=parser.parse_args() |
| source=Path(args.source).resolve();out=Path(args.output).resolve();assert not out.exists();out.mkdir(parents=True);torch.cuda.set_device(args.device);start=time.perf_counter();trees=[];tensor_count=0 |
| for name in ['run_01','objective_extension_01']: |
| directory=out/name;directory.mkdir() |
| for file in sorted((source/name).iterdir()): |
| if file.suffix=='.pt': |
| original=torch.load(file,map_location=args.device,weights_only=True);base=directory/file.stem;save_tree(original,base) |
| restored=load_tree(base,args.device);count=compare(original,restored);tensor_count+=count |
| trees.append(dict(file=str(base.relative_to(out)),source_sha256=sha(file),tensors=count,json_sha256=sha(base.with_suffix('.json')),safetensors_sha256=sha(base.with_suffix('.safetensors')))) |
| elif file.suffix in ('.json','.csv','.md','.py'):shutil.copyfile(file,directory/file.name) |
| for name in ['analysis_spec.json','objective_extension_spec.json','findings.json','findings.md','primary_table.csv','full_replay.json','verify_all.py','safe_tree.py','objective_extension.py','export_safe.py']: |
| shutil.copyfile(source/name,out/name) |
| trainer=source.parents[2]/'reproduce/review_r1_sensor.py';shutil.copyfile(trainer,out/'review_r1_sensor.py') |
| assert not list(out.rglob('*.pt')) |
| report=dict(format='Lossless Safetensors + JSON tensor trees; no executable pickle checkpoints',trees=trees,roundtrip_tensor_count=tensor_count,roundtrip_maximum_difference=0,seconds=time.perf_counter()-start, |
| source_artifacts_modified=False,device=args.device,export_uses_cpu_only_for_serialization=True, |
| replay_command='python verify_all.py --release /path/to/original-release --base run_01 --extension objective_extension_01 --output replay.json --device cuda:0') |
| (out/'EXPORT_MANIFEST.json').write_text(json.dumps(report,indent=2)+'\n');print(json.dumps({k:v for k,v in report.items() if k!='trees'},indent=2)) |
|
|
|
|
| if __name__=='__main__':main() |
|
|
