robottt-pi05-base-comparison / code /tests /test_config_storage.py
jisue's picture
Experiment backup 2026-11-11T03:59:02.166535+00:00
a34b47d verified
Raw History Blame Contribute Delete
4.68 kB
import copy,json,tempfile,unittest
from pathlib import Path
import torch
from robottt_pi05.config import ExperimentConfig,ADAMW_SETTINGS
from robottt_pi05.modeling import extension,configure_trainables
from robottt_pi05.smoke_optimizer import build_optimizer
from robottt_pi05.checkpoint import save_boundary,restore_boundary,load_weights,digest,capture_rng,resume_step
from robottt_pi05.training_io import RunLedger,finish_output
from robottt_pi05.eval_io import atomic_tensor,atomic_json,completed_output,sha
class ConfigStorageTests(unittest.TestCase):
def test_modes_and_mapping(self):
for cfg,count,objects in [(ExperimentConfig(active_layers=(17,)),152672,10),(ExperimentConfig(),561536,37),(ExperimentConfig(learn_inner_lr=True),561540,41),(ExperimentConfig(active_layers=tuple(range(18))),2469568,163),(ExperimentConfig(active_layers=tuple(range(18)),learn_inner_lr=True),2469586,181)]:
module=extension(cfg);names,params=configure_trainables(module,cfg)
self.assertEqual((sum(p.numel() for p in params),len(params)),(count,objects))
self.assertEqual(sum(n.endswith('.raw_lr') for n in names),len(cfg.active_layers) if cfg.learn_inner_lr else 0)
for kw in [dict(active_layers=(17,4)),dict(segment_length=8),dict(full_policy_training=True),dict(learn_inner_lr=True,segment_length=16),dict(state_policy='none')]:
with self.assertRaises((ValueError,NotImplementedError)):ExperimentConfig(**kw)
def test_checkpoint_actual_optimizer_roundtrip(self):
cfg=ExperimentConfig(learn_inner_lr=True);module=extension(cfg);names,params=configure_trainables(module,cfg)
optimizer=build_optimizer(params,ADAMW_SETTINGS)
# CPU synthetic optimizer unit test. No VLA, policy forward or research input.
sum(p.square().sum() for p in params).backward();optimizer.step();optimizer.zero_grad(set_to_none=True)
expected=copy.deepcopy(module.state_dict());moments=digest(optimizer.state_dict());rng=capture_rng()
with tempfile.TemporaryDirectory() as tmp:
path,data=save_boundary(tmp,step=1,slow=module,optimizer=optimizer,config=cfg,provenance={'fixture':'synthetic'},attempted={'optimizer_update':1},completed={'optimizer_update':1})
with torch.no_grad():
for p in params:p.add_(1)
torch.rand(3)
restored=restore_boundary(path,slow=module,optimizer=optimizer,config=cfg,provenance={'fixture':'synthetic'})
self.assertEqual(restored['next_schedule_index'],1);self.assertEqual(digest(optimizer.state_dict()),moments);self.assertEqual(digest(capture_rng()),digest(rng))
for name,value in module.state_dict().items():self.assertTrue(torch.equal(value,expected[name]))
raw=[x.raw_lr.clone() for x in module.ttt.values()]
configure_trainables(module,cfg,inference=True);module.eval();module.train();configure_trainables(module,cfg)
for a,b in zip(raw,module.ttt.values()):self.assertTrue(torch.equal(a,b.raw_lr))
bad=copy.deepcopy(data);del bad['slow']['ttt.4.raw_lr'];bp=Path(tmp)/'bad.pt';atomic_tensor(bp,bad)
with self.assertRaises(ValueError):load_weights(bp,module,cfg,expected_sha256=sha(bp))
bad=copy.deepcopy(data);bad['active_layers']=[17];bp=Path(tmp)/'bad2.pt';atomic_tensor(bp,bad)
with self.assertRaises(ValueError):load_weights(bp,module,cfg,expected_sha256=sha(bp))
def test_durable_profile_failure_and_no_duplicate(self):
with tempfile.TemporaryDirectory() as tmp:
d=Path(tmp);ledger=RunLedger(d,{'optimizer_update':2,'case':1});path=d/'case.pt'
ledger.reserve('case','once');atomic_tensor(path,{'key':'once','synthetic':True});ledger.complete('case','once')
with self.assertRaises(ValueError):finish_output(ledger=ledger,key='once',path=path,profile={'scope':'collision'})
reopened=RunLedger(d,ledger.limits)
self.assertTrue(completed_output(reopened,'once',path));self.assertEqual(reopened.attempted['case'],1)
with self.assertRaises(ValueError):reopened.profile(scope='x',values={'x':float('nan')})
with self.assertRaises(ValueError):reopened.event('x',epoch=1)
partial=d/'unfinished.pt.partial.1';partial.write_bytes(b'not a completed tensor')
self.assertFalse(completed_output(reopened,'unfinished',partial))
reopened.reserve('optimizer_update','1')
with self.assertRaises(RuntimeError):resume_step(d/'checkpoints',reopened)
self.assertEqual(reopened.attempted['optimizer_update'],1)
if __name__=='__main__':unittest.main()