transform.py

1from pathlib import Path
2H=Path(__file__).resolve().parent
3p=H/'train.py';s=p.read_text()
4s=s.replace("'source_updates':250,'source_cursor':128000", "'source_updates':saved['updates'],'source_cursor':saved['exposures']['cursor']")
5s=s.replace("attempt='recovery'", "attempt='completion'")
6s=s.replace("best=latest['nll'];best_validation=latest", "best=saved['best_nll'];best_validation=saved['best_validation']")
7s=s.replace("'cm7-theory-recovery-v1'", "'cm7-theory-completion-v1'")
8a=s.index('        # Certify the exact restored update250');b=s.index('        dist.barrier()',a)
9s=s[:a]+'''        # Carry the previous loss-selected best independently of last-state resume.
10        if rank==0:
11            prior=torch.load(cfg['prior_selected']['path'],map_location='cpu',weights_only=True)
12            if sha(cfg['prior_selected']['path'])!=cfg['prior_selected']['sha256'] or prior['latest_validation']['nll']!=best:raise ValueError('prior selected loss/hash')
13            prior.update(format='cm7-theory-completion-v1',identity=cfg,identity_sha256=digest(cfg),deadline=deadline_value,is_best_snapshot=True,terminal=False,attempt='inherited-loss-best')
14            capacity(cfg);atomic_save(out/'best-inherited.pt',prior);del prior
15        dist.barrier()
16        ref=[{'path':str(out/'best-inherited.pt'),'sha256':sha(out/'best-inherited.pt')} if rank==0 else None];dist.broadcast_object_list(ref,0);best_ref=baseline_ref=ref[0]
17        if rank==0:
18            from finalization import atomic
19            atomic(out/'baseline.json',{'checkpoint':baseline_ref,'config_sha256':digest(cfg)})
20            atomic_json(out/'best.json',best_ref);atomic_json(out/'initial-validation.json',initial_validation)
21            atomic_json(out/'resume-receipt.json',{'updates':updates,'microstep':cursor,'cursor':plan.cursor(cursor),'optimizer':opt.retention_proof,'source_checkpoint_sha256':cfg['source_checkpoint_sha256'],'authorized_completion_deadline':deadline_value,'prior_selected':cfg['prior_selected']})
22''' +s[b:]
23p.write_text(s)
24p=H/'resume_state.py';s=p.read_text();a=s.index('def validate_source(');b=s.index('def restore_optimizer_scheduler',a)
25s=s[:a]+'''def validate_source(saved,cfg,clock,ledger,plan):
26    original=json.loads(Path(cfg['original_config']).read_text())
27    if sha(cfg['original_config'])!=cfg['original_config_sha256'] or saved['identity']!=original or saved['identity_sha256']!=digest(original):raise ValueError('source identity')
28    if saved['format']!='cm7-theory-recovery-v1' or not saved['terminal'] or saved['retention_failures']:raise ValueError('clean loss-only terminal source required')
29    if saved['deadline']!=json.loads(Path(cfg['source_deadline']).read_text()):raise ValueError('source clock changed')
30    consumed=ledger.verify(saved['exposures'])
31    if consumed!=cfg['source_cursor'] or saved['updates']!=cfg['source_updates'] or saved['microstep']!=cfg['source_microstep'] or plan.cursor(saved['microstep'])!=consumed:raise ValueError('source sampler/counter mismatch')
32    if consumed>=len(ledger.data.schedule):raise ValueError('finite schedule already complete')
33    if len(saved['rank_rng'])!=2 or len(saved['rank_group_totals'])!=2 or sum(float(t[:,0].sum()) for t in saved['rank_group_totals'])!=consumed:raise ValueError('source rank state')
34    if saved['scheduler_state_dict']['last_epoch']!=saved['updates']:raise ValueError('source scheduler step')
35    for state in saved['optimizer_state_dict']['state'].values():
36        if float(state['step'])!=19000+saved['updates'] or not torch.isfinite(state['exp_avg']).all() or not torch.isfinite(state['exp_avg_sq']).all():raise ValueError('source Adam state')
37    if saved['best_checkpoint']!=cfg['prior_selected']:raise ValueError('prior selected reference changed')
38    return consumed
39
40''' +s[b:]
41p.write_text(s)
42