File size: 20,309 Bytes
2d5c26a 1a0a7fb 2d5c26a 1a0a7fb 2d5c26a 1a0a7fb 2d5c26a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 | import copy
from datetime import datetime, timezone
import hashlib
import json
import io
import math
from pathlib import Path
import tempfile
import time
import unittest
from unittest.mock import patch, MagicMock
from selection import SELECTION_METRIC, validation_selection
from scripts.fleet_campaign import (choose_candidate, checksum, copy_evidence, identity,
snapshot_candidate, validate_plan, validation_score,
verify_candidate, verify_correctness, verify_compatibility,
verify_candidate_dataset, verify_selected_dataset, PROTECTED_SPLITS,
deploy, stop_candidate, ROOT)
def predictions(logit=1.0):
normalizer = math.log1p(math.exp(-logit))
logs = [-normalizer, -logit-normalizer]
return [{'id':f'{family}-{index}','group':f'{family}-{index}','family':family,
'choices':['yes','no'],'target':0,'logits':[logit,0.0],
'log_probabilities':logs[:],'probabilities':[math.exp(v) for v in logs]}
for family in ('arc','banking','boolq','snli') for index in range(128)]
def artifact(rows, model):
from training_model import ADAPTER_VERSION, PROMPT_VERSION
dataset = Path(model).parent/'dataset'
dataset.mkdir(exist_ok=True)
(dataset/'validation.jsonl').write_text('frozen opaque bytes')
hashes = {'validation':checksum(dataset/'validation.jsonl')}
(dataset/'manifest.json').write_text(json.dumps({'split_sha256':hashes}))
signature = hashlib.sha256(json.dumps({'data':hashes,'model':{'revision':'pinned'},
'implementation':checksum(ROOT/'training_model.py'),'max_tokens':512},sort_keys=True).encode()).hexdigest()
return {'format':'opensysone-adapter-v1','step':40,'trainable_state':{'head':'fake CPU test tensor'},
'best_validation_macro_nll':validation_score(rows,rows)['macro_nll'],
'selection_metric':SELECTION_METRIC,'best_validation_selection_score':validation_selection(rows)['score'],
'source_commit':'abc123','model_provenance':{'revision':'pinned'},'data_signature':signature,
'adapter_version':ADAPTER_VERSION,'prompt_version':PROMPT_VERSION,
'config':{'validation_per_family':128,'model':str(model),'dataset':str(dataset),'max_tokens':512}}
def correctness(directory):
checks = {key:0.0 for key in ('branch_chunks_1_probability_max_abs','branch_chunks_2_probability_max_abs',
'branch_chunks_4_probability_max_abs','question_isolation_probability_max_abs',
'candidate_permutation_probability_max_abs','repeat_probability_max_abs')}
checks['tolerance_probability_abs'] = 1e-4
(directory/'correctness_initial.json').write_text(json.dumps(checks))
return checks
def dataset_variant(directory, train='original training'):
directory.mkdir()
hashes = {}
for split in ('train', *PROTECTED_SPLITS):
path = directory/f'{split}.jsonl'
path.write_text(train if split == 'train' else 'opaque frozen ' + split)
hashes[split] = checksum(path)
(directory/'manifest.json').write_text(json.dumps({'split_sha256':hashes}))
return directory
class FleetCampaignTests(unittest.TestCase):
def test_dataset_override_requires_explicit_matching_registration_and_protected_bytes(self):
with tempfile.TemporaryDirectory() as directory:
root = Path(directory)
baseline = dataset_variant(root/'v1')
expanded = dataset_variant(root/'v2', 'original plus additional training')
candidate = {'name':'expanded','host':'local','dataset':str(expanded),
'project':'/project','python':'/python','campaign':'/campaign','training':'/training'}
plan = {'dataset':str(baseline),'candidates':[candidate], 'reference_predictions':'/reference.json',
'selection_cutoff':'2026-09-17T16:00:00Z','final_deadline':'2026-09-17T18:16:10Z'}
proof = verify_candidate_dataset(candidate, plan, expanded)
self.assertEqual(set(proof['protected_split_sha256']),set(PROTECTED_SPLITS))
self.assertNotEqual(checksum(baseline/'train.jsonl'),checksum(expanded/'train.jsonl'))
self.assertLess(*validate_plan(plan, datetime(2026,9,17,6,tzinfo=timezone.utc).timestamp()))
selected = {'name':'expanded','config':{'dataset':str(expanded)}}
self.assertEqual(verify_selected_dataset(selected,plan),proof)
without_override = {key:value for key,value in candidate.items() if key != 'dataset'}
self.assertIsNone(verify_candidate_dataset(without_override,plan,baseline))
for entry,actual in ((without_override,expanded),(candidate,baseline),
({**candidate,'dataset':'relative'},expanded)):
with self.assertRaises(ValueError):
verify_candidate_dataset(entry,plan,actual)
with self.assertRaisesRegex(ValueError,'not uniquely registered'):
verify_selected_dataset({**selected,'name':'unregistered'},plan)
def test_dataset_override_rejects_each_changed_missing_or_unhashed_protected_split(self):
for split in PROTECTED_SPLITS:
for damage in ('changed','missing_checksum','missing_file','stale_checksum'):
with self.subTest(split=split,damage=damage), tempfile.TemporaryDirectory() as directory:
root = Path(directory)
baseline = dataset_variant(root/'v1')
expanded = dataset_variant(root/'v2','expanded training')
manifest = json.loads((expanded/'manifest.json').read_text())
if damage in ('changed','stale_checksum'):
(expanded/f'{split}.jsonl').write_text('different reserved bytes')
if damage == 'changed':
manifest['split_sha256'][split] = checksum(expanded/f'{split}.jsonl')
elif damage == 'missing_checksum':
del manifest['split_sha256'][split]
else:
(expanded/f'{split}.jsonl').unlink()
(expanded/'manifest.json').write_text(json.dumps(manifest))
with self.assertRaisesRegex(ValueError,'[Pp]rotected split'):
verify_candidate_dataset({'dataset':str(expanded)}, {'dataset':str(baseline)}, expanded)
def test_fresh_finalization_uses_verified_selected_dataset_variant(self):
with tempfile.TemporaryDirectory() as directory:
root = Path(directory)
baseline = dataset_variant(root/'v1')
expanded = dataset_variant(root/'v2','expanded training')
campaign = root/'campaign'
campaign.mkdir()
selected = {'name':'expanded','checkpoint':'/verified/best.pt','config':{'dataset':str(expanded)}}
plan = {'dataset':str(baseline),'candidates':[{'name':'expanded','dataset':str(expanded)}]}
verify_selected_dataset(selected,plan)
deadline = time.time()+7200
with patch('scripts.fleet_campaign.stage',side_effect=RuntimeError('CPU fixture stops before execution')) as stage:
with self.assertRaisesRegex(RuntimeError,'CPU fixture'):
deploy(campaign,{},plan,selected,deadline)
self.assertEqual(stage.call_args.args[2],'evaluation')
command = stage.call_args.args[3]
self.assertEqual(command[command.index('--dataset')+1],str(expanded))
self.assertEqual(stage.call_args.args[4],deadline-120)
def test_natural_exit_between_inspection_and_stop_requires_fresh_stopped_status(self):
running = {'supervisor_running':True,'child_running':True,'api_running':False}
stopped = {'supervisor_running':False,'child_running':False,'api_running':False}
for followup in (stopped,running):
with self.subTest(followup=followup), patch('scripts.fleet_campaign.candidate_status',
side_effect=[running,RuntimeError('No recorded process remains'),followup]) as status:
if followup is stopped:
self.assertEqual(stop_candidate({'name':'gx10'}),stopped)
else:
with self.assertRaisesRegex(RuntimeError,'No recorded process'):
stop_candidate({'name':'gx10'})
self.assertEqual(status.call_count,3)
self.assertEqual(status.call_args_list[1].kwargs,{'stop':True})
self.assertEqual(status.call_args_list[2].kwargs,{})
def test_validation_identity_probabilities_and_durable_score_are_required(self):
rows = predictions()
with tempfile.TemporaryDirectory() as directory:
saved = artifact(rows,Path(directory)/'model')
self.assertAlmostEqual(verify_candidate(saved,rows,rows)['macro_nll'],math.log1p(math.exp(-1)))
self.assertEqual(verify_candidate(saved,list(reversed(rows)),rows)['accuracy'],1)
changed = copy.deepcopy(rows)
changed[0]['id'] = 'different-id'
with self.assertRaisesRegex(ValueError,'identical 512'):
verify_candidate(saved,changed,rows)
changed = copy.deepcopy(rows)
changed[0]['group'] = 'different-source-group'
with self.assertRaises(ValueError):
verify_candidate(saved,changed,rows)
changed = copy.deepcopy(rows)
changed[0]['probabilities'] = [0.5,0.5]
with self.assertRaisesRegex(ValueError,'disagree with logits'):
verify_candidate(saved,changed,rows)
changed = copy.deepcopy(rows)
changed[0]['logits'][0] = float('nan')
with self.assertRaisesRegex(ValueError,'Non-finite'):
verify_candidate(saved,changed,rows)
with self.assertRaisesRegex(ValueError,'durable best'):
verify_candidate({**saved,'best_validation_macro_nll':0.01},rows,rows)
with self.assertRaisesRegex(ValueError,'durable trained artifact'):
verify_candidate({**saved,'trainable_state':{}},rows,rows)
with self.assertRaisesRegex(ValueError,'raw validation'):
verify_candidate({**saved,'temperature':2},rows,rows)
with self.assertRaisesRegex(ValueError,'selection policy'):
verify_candidate({**saved,'selection_metric':'raw_nll'},rows,rows)
with self.assertRaisesRegex(ValueError,'Crossfit selection score'):
verify_candidate({**saved,'best_validation_selection_score':0.2},rows,rows)
with self.assertRaises(ValueError):
verify_candidate(saved,rows[:-1],rows)
duplicate = copy.deepcopy(rows)
duplicate[-1] = duplicate[0]
with self.assertRaisesRegex(ValueError,'Duplicate'):
identity(duplicate)
def test_selection_ignores_incomplete_better_scores(self):
candidates = [{'name':'incomplete','eligible':False,'metrics':{'selection_score':0.01}},
{'name':'gx10','eligible':True,'metrics':{'macro_nll':0.3,'selection_score':0.4}},
{'name':'spark','eligible':True,'metrics':{'macro_nll':0.5,'selection_score':0.3}}]
self.assertEqual(choose_candidate(candidates)['name'],'spark')
self.assertEqual(choose_candidate(candidates[:2])['name'],'gx10')
with self.assertRaisesRegex(RuntimeError,'No verified durable'):
choose_candidate(candidates[:1])
def test_deadlines_timezone_paths_and_fallback(self):
now = datetime(2026,9,16,20,tzinfo=timezone.utc).timestamp()
plan = {'selection_cutoff':'2026-09-17T16:00:00Z','final_deadline':'2026-09-17T18:16:10Z',
'dataset':'/frozen/data','reference_predictions':'/reference.json',
'candidates':[{'name':'gx10','host':'local','project':'/project','python':'/python',
'campaign':'/campaign','training':'/campaign/training'}]}
self.assertLess(*validate_plan(plan,now))
for change in ({'final_deadline':'2026-09-17T17:00:00Z'},
{'final_deadline':'2026-09-17T18:16:10'},
{'final_deadline':'2026-09-18T18:16:10Z'},
{'reference_predictions':'relative'},
{'candidates':[{**plan['candidates'][0],'host':'192.168.8.111'}]}):
with self.assertRaises(ValueError):
validate_plan({**plan,**change},now)
def test_snapshot_loads_durable_best_and_never_copies_test_predictions(self):
import torch
with tempfile.TemporaryDirectory() as directory:
root = Path(directory)
training = root/'training'
training.mkdir()
model = root/'model'
model.mkdir()
rows = predictions()
saved = artifact(rows,model)
(model/'opensysone-provenance.json').write_text(json.dumps(saved['model_provenance']))
torch.save(saved,training/'best.pt')
correctness(training)
(training/'best_validation_predictions.json').write_text(json.dumps(rows))
(training/'test_predictions.json').write_text('MUST NOT COPY')
(training/'calibration_predictions.json').write_text('MUST NOT COPY')
(training/'holdout_predictions.json').write_text('MUST NOT COPY')
(training/'checkpoint.pt').write_text('MUST NOT COPY OPTIMIZER')
candidate = {'name':'gx10','host':'local','training':str(training),'campaign':'/campaign'}
record = snapshot_candidate(candidate,root/'snapshot',rows)
self.assertTrue(record['eligible'])
self.assertEqual(record['checkpoint_sha256'],checksum(training/'best.pt'))
copied = {p.name for p in (root/'snapshot/training').iterdir()}
self.assertEqual(copied,{'best.pt','best_validation_predictions.json','correctness_initial.json'})
(training/'best_validation_predictions.json').unlink()
with self.assertRaisesRegex(ValueError,'no matching complete validation'):
snapshot_candidate(candidate,root/'missing-evidence',rows)
def test_snapshot_accepts_inherited_pilot_evidence_and_rejects_transfer_change(self):
import torch
with tempfile.TemporaryDirectory() as directory:
root = Path(directory)
training,pilot,model = root/'training',root/'pilot',root/'model'
for path in (training,pilot,model):
path.mkdir()
rows = predictions()
torch.save(artifact(rows,model),training/'best.pt')
(model/'opensysone-provenance.json').write_text(json.dumps({'revision':'pinned'}))
correctness(training)
(pilot/'validation_step_000040_predictions.json').write_text(json.dumps(rows))
candidate = {'name':'gx10','host':'local','training':str(training),'campaign':'/campaign',
'evidence_dirs':[str(pilot)]}
record = snapshot_candidate(candidate,root/'snapshot',rows)
self.assertTrue(record['eligible'])
with patch('scripts.fleet_campaign.command_on',side_effect=['first-sha best.pt','changed-sha best.pt']):
with self.assertRaisesRegex(ValueError,'changed during snapshot'):
snapshot_candidate(candidate,root/'changed-snapshot',rows)
def test_correctness_requires_complete_finite_gate_and_does_not_hide_final_failure(self):
with tempfile.TemporaryDirectory() as directory:
directory = Path(directory)
with self.assertRaisesRegex(ValueError,'lacks'):
verify_correctness(directory)
checks = correctness(directory)
self.assertEqual(verify_correctness(directory).name,'correctness_initial.json')
checks['repeat_probability_max_abs'] = float('nan')
(directory/'correctness_final.json').write_text(json.dumps(checks))
with self.assertRaisesRegex(ValueError,'failed'):
verify_correctness(directory)
checks['repeat_probability_max_abs'] = 0.0
checks['tolerance_probability_abs'] = 0.02
(directory/'correctness_final.json').write_text(json.dumps(checks))
with self.assertRaisesRegex(ValueError,'weakened'):
verify_correctness(directory)
def test_compatibility_checks_versions_and_actual_frozen_source_bytes(self):
with tempfile.TemporaryDirectory() as directory:
saved = artifact(predictions(),Path(directory)/'model')
verify_compatibility(saved)
for change in ({'prompt_version':'other'}, {'adapter_version':'other'}, {'data_signature':'other'}):
with self.assertRaisesRegex(ValueError,'incompatible'):
verify_compatibility({**saved,**change})
(Path(saved['config']['dataset'])/'validation.jsonl').write_text('modified')
with self.assertRaisesRegex(ValueError,'source hash mismatch'):
verify_compatibility(saved)
def test_deployment_pointer_requires_completed_eval_and_actual_http_probe(self):
for succeeds in (False,True):
with self.subTest(succeeds=succeeds), tempfile.TemporaryDirectory() as directory:
root = Path(directory)
campaign = root/'campaign'
evaluation = campaign/'evaluation'
evaluation.mkdir(parents=True)
model = evaluation/'model.pt'
model.write_bytes(b'durable final artifact')
(evaluation/'metrics.json').write_text(json.dumps({'status':'complete','model_sha256':checksum(model)}))
(evaluation/'manifest.json').write_text(json.dumps({'checkpoint_sha256':'selected-sha',
'data_signature':'frozen-data','source_sha256':{name:checksum(ROOT/name)
for name in ('experiment.py','training_model.py')}}))
selected = {'checkpoint_sha256':'selected-sha','data_signature':'frozen-data'}
api = MagicMock(pid=987654321)
api.poll.return_value = None
responses = [io.BytesIO(json.dumps({'status':'ready','checkpoint':str(model),'model':'opensysone'}).encode()),
io.BytesIO(b'{"response":"HTTP probe"}')]
def validate(payload,result):
self.assertFalse((root/'ai/opensysone/deploy/current.json').exists())
self.assertEqual(result,{'response':'HTTP probe'})
if not succeeds:
raise ValueError('invalid actual HTTP response')
state = {}
with patch('scripts.fleet_campaign.runner.STOP',False), \
patch('scripts.fleet_campaign.stage') as stage_mock, \
patch('scripts.fleet_campaign.subprocess.Popen',return_value=api), \
patch('scripts.fleet_campaign.urllib.request.urlopen',side_effect=responses) as http, \
patch('scripts.fleet_campaign.Path.home',return_value=root), \
patch('jev_harness.validate_response',side_effect=validate):
if succeeds:
deploy(campaign,state,{'port':18081},selected,time.time()+120)
pointer = json.loads((root/'ai/opensysone/deploy/current.json').read_text())
self.assertEqual(pointer['sha256'],checksum(model))
self.assertEqual(state['status'],'complete')
api.terminate.assert_not_called()
else:
with self.assertRaisesRegex(ValueError,'invalid actual HTTP'):
deploy(campaign,state,{'port':18081},selected,time.time()+120)
self.assertFalse((root/'ai/opensysone/deploy/current.json').exists())
api.terminate.assert_called_once()
self.assertEqual(stage_mock.call_args.args[2],'harness_check')
self.assertEqual(http.call_args.args[0].method,'POST')
if __name__ == '__main__':
unittest.main()
|