File size: 21,154 Bytes
1a0a7fb | 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 | """Prepare an immutable expanded-data pilot backup; never upload or load a model.
Usage: python scripts/prepare_expansion_backup.py --pilot /path/to/completed-run
--output /path/to/new-export --fleet-plan /path/to/fleet/plan.json
The output is accepted by scripts/publish_hf_snapshot.py. Publication is separate.
Only trusted local project checkpoints may be supplied (PyTorch pickle format).
"""
import argparse
from datetime import datetime, timezone
import hashlib
import json
from pathlib import Path
import re
import shutil
import subprocess
import sys
import tempfile
PROJECT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(PROJECT))
from scripts.fleet_campaign import verify_candidate, verify_correctness
from scripts.publish_hf_snapshot import relative_name, verify_artifacts
SPLITS = ('train', 'validation', 'calibration', 'test', 'holdout')
COMPLETE = {'completed_step_target', 'completed_epochs', 'validation_early_stop'}
def digest(path):
result = hashlib.sha256()
with Path(path).open('rb') as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b''):
result.update(chunk)
return result.hexdigest()
def read_json(path):
return json.loads(Path(path).read_text())
def write_json(path, value):
Path(path).write_text(json.dumps(value, indent=2, sort_keys=True) + '\n')
def verified_dataset(path):
value = read_json(path / 'manifest.json')
if set(value.get('split_sha256', {})) != set(SPLITS):
raise ValueError('Dataset must contain exactly five registered splits')
for split, checksum in value['split_sha256'].items():
if digest(path / (split + '.jsonl')) != checksum:
raise ValueError('Dataset split checksum mismatch')
return value
def verify_source(root, manifest):
revision = manifest['git_commit']
if not re.fullmatch('[0-9a-f]{40}', revision):
raise ValueError('Training source revision is not exact')
if not manifest.get('source_sha256'):
raise ValueError('Training source hash inventory is missing')
for name, expected in manifest['source_sha256'].items():
relative_name(name)
content = subprocess.check_output(['git', 'show', revision + ':' + name], cwd=root,
stderr=subprocess.DEVNULL, timeout=30)
if hashlib.sha256(content).hexdigest() != expected:
raise ValueError('Training source does not match committed revision')
return revision
def finite_tensors(value, torch):
if isinstance(value, torch.Tensor):
return value.device.type == 'cpu' and bool(torch.isfinite(value).all())
if isinstance(value, dict):
return all(finite_tensors(item, torch) for item in value.values())
if isinstance(value, (list, tuple)):
return all(finite_tensors(item, torch) for item in value)
return True
def prepare(pilot, output, fleet_plan, repo_id='andyshu/opensysone', source_root=PROJECT):
import torch
torch.set_num_threads(4)
if torch.cuda.is_initialized():
raise ValueError('Backup must run without initialized CUDA')
pilot, output, source_root = Path(pilot).resolve(), Path(output).resolve(), Path(source_root).resolve()
if (pilot / 'artifacts').is_dir():
pilot = pilot / 'artifacts'
if output.exists() or output.is_relative_to(source_root) or output == source_root:
raise ValueError('Output must be a new directory outside source checkout')
if (pilot.parent / 'exit_code').read_text().strip() != '0':
raise ValueError('Pilot must have completed with exit code zero')
manifest, summary = read_json(pilot / 'manifest.json'), read_json(pilot / 'summary.json')
if summary.get('status') not in COMPLETE or summary.get('final_correctness_status') != 'passed':
raise ValueError('Pilot is not complete with final correctness gates passed')
verify_correctness(pilot)
revision = verify_source(source_root, manifest)
selected_hash, checkpoint_hash = digest(pilot / 'best.pt'), digest(pilot / 'checkpoint.pt')
if summary.get('best_sha256') != selected_hash or summary.get('checkpoint_sha256') != checkpoint_hash:
raise ValueError('Pilot summary does not match durable checkpoints')
best = torch.load(pilot / 'best.pt', map_location='cpu', weights_only=False)
latest = torch.load(pilot / 'checkpoint.pt', map_location='cpu', weights_only=False)
for saved in (best, latest):
if saved.get('format') != 'opensysone-adapter-v1' or not saved.get('trainable_state') or 'temperature' in saved:
raise ValueError('Expected raw trained adapter/head checkpoint')
if saved.get('source_commit') != revision or saved.get('config') != manifest['config']:
raise ValueError('Checkpoint source/config differs from pilot manifest')
if saved.get('data_signature') != manifest['data_signature'] or saved.get('model_provenance') != manifest['model_provenance']:
raise ValueError('Checkpoint data/model provenance differs from pilot manifest')
if not finite_tensors(saved.get('trainable_state'), torch):
raise ValueError('Checkpoint trainable tensors are nonfinite or non-CPU')
if latest['step'] != summary['completed_steps'] or not 0 <= best['step'] <= latest['step']:
raise ValueError('Checkpoint steps disagree with completed pilot')
for key in ('best_validation_macro_nll', 'best_validation_selection_score', 'selection_metric'):
if latest.get(key) != best.get(key) or summary.get(key) != best.get(key):
raise ValueError('Resumable checkpoint does not preserve selected best metrics')
if not {'optimizer', 'random_state', 'torch_rng', 'cuda_rng'}.issubset(latest):
raise ValueError('Resumable checkpoint lacks optimizer or random states')
optimizer = latest['optimizer']
if not optimizer.get('state') or not finite_tensors(optimizer, torch):
raise ValueError('Resumable optimizer state is missing or nonfinite')
optimizer_steps = sorted({float(item['step']) for item in optimizer['state'].values()})
if optimizer_steps != [float(latest['step'])] or latest['step'] <= 0:
raise ValueError('Optimizer step does not match the completed pilot')
dataset = Path(best['config']['dataset']).resolve()
data = verified_dataset(dataset)
base_root = Path(data['base_dataset']['path']).resolve()
base = verified_dataset(base_root)
if (data['base_dataset']['manifest_sha256'] != digest(base_root / 'manifest.json') or
data['base_dataset']['split_sha256'] != base['split_sha256'] or
any(data['split_sha256'][s] != base['split_sha256'][s] for s in SPLITS[1:])):
raise ValueError('Expanded dataset parent/protected split lineage mismatch')
def signature(value):
return hashlib.sha256(json.dumps({'data': value['split_sha256'], 'model': best['model_provenance'],
'implementation': manifest['source_sha256']['training_model.py'],
'max_tokens': best['config']['max_tokens']}, sort_keys=True).encode()).hexdigest()
expected_transition = {'kind': 'training_split_only', 'parent_dataset': str(base_root),
'dataset': str(dataset), 'parent_data_signature': signature(base), 'data_signature': signature(data),
'parent_manifest_sha256': digest(base_root / 'manifest.json'),
'manifest_sha256': digest(dataset / 'manifest.json'),
'protected_split_sha256': {s: base['split_sha256'][s] for s in SPLITS[1:]}}
for saved in (best, latest):
initial = saved.get('initialization', {})
if (initial.get('kind') != 'warm_start' or initial.get('restores_optimizer') is not False or
initial.get('restores_rng') is not False or initial.get('data_transition') != expected_transition or
saved['data_signature'] != signature(data)):
raise ValueError('Expanded pilot lacks verified fresh-optimizer data transition')
if data['source_script_sha256'] != manifest['source_sha256']['scripts/prepare_expanded_data.py']:
raise ValueError('Expanded dataset builder differs from committed training source')
diagnostics = data['diagnostics']
diagnostic_name = relative_name(diagnostics['path'])
if diagnostics.get('selection_eligible') is not False or digest(dataset / diagnostic_name) != diagnostics['sha256']:
raise ValueError('Diagnostic data checksum or selection exclusion mismatch')
plan_hash = digest(fleet_plan)
plan = read_json(fleet_plan)
plan_captured_utc = datetime.now(timezone.utc).isoformat()
reference_path = Path(plan['reference_predictions'])
if digest(reference_path) != plan['reference_sha256']:
raise ValueError('Fixed fleet validation reference checksum mismatch')
reference = read_json(reference_path)
parent_path = Path(best['initialization']['parent_checkpoint'])
parent_hash = best['initialization']['parent_checkpoint_sha256']
if digest(parent_path) != parent_hash:
raise ValueError('Warm-start parent checkpoint checksum mismatch')
parent = torch.load(parent_path, map_location='cpu', weights_only=False)
parent_manifest = read_json(parent_path.parent / 'manifest.json')
parent_revision = verify_source(source_root, parent_manifest)
if (parent['source_commit'] != parent_revision or parent_revision != best['initialization']['parent_source_commit'] or
parent['step'] != best['initialization']['parent_step'] or parent['data_signature'] != signature(base) or
parent['model_provenance'] != best['model_provenance'] or not finite_tensors(parent['trainable_state'], torch)):
raise ValueError('Warm-start parent provenance mismatch')
verify_correctness(parent_path.parent)
parent_predictions = parent_path.parent / f"validation_step_{parent['step']:06d}_predictions.json"
parent_metrics = verify_candidate(parent, read_json(parent_predictions), reference)
if read_json(parent_path.parent / 'best_validation_selection.json') != parent_metrics['selection']:
raise ValueError('Parent selection evidence disagrees with durable checkpoint')
predictions_path = None
choices = [pilot / f"validation_step_{best['step']:06d}_predictions.json",
pilot / 'best_validation_predictions.json']
if best['step'] == 0:
choices.append(pilot / 'initial_validation_predictions.json')
for path in choices:
if path.is_file():
try:
metrics = verify_candidate(best, read_json(path), reference)
except ValueError:
continue
predictions_path = path
break
if predictions_path is None:
raise ValueError('No matching fixed-validation evidence for durable best')
if read_json(pilot / 'best_validation_selection.json') != metrics['selection']:
raise ValueError('Pilot selection evidence disagrees with durable checkpoint')
output.parent.mkdir(parents=True, exist_ok=True)
staging = Path(tempfile.mkdtemp(prefix=output.name + '.staging-', dir=output.parent))
files = {}
def copy(source, name, expected=None):
source = Path(source)
name = relative_name(name)
before = digest(source)
if expected is not None and expected != before:
raise ValueError('Source checksum differs from recorded provenance')
target = staging / name
target.parent.mkdir(parents=True, exist_ok=True)
shutil.copyfile(source, target)
if digest(source) != before or digest(target) != before:
raise ValueError('Source changed while creating immutable backup')
files[name] = {'bytes': target.stat().st_size, 'sha256': before, 'source_path': str(source)}
artifact_directory = 'artifacts/expanded-gx10-4b-pilot'
try:
copy(fleet_plan, 'fleet/plan.json', plan_hash)
copy(reference_path, 'fleet/reference-validation-predictions.json', plan['reference_sha256'])
copy(Path(__file__).resolve(), 'backup-tools/prepare_expansion_backup.py')
test_helper = PROJECT / 'tests/test_expansion_backup.py'
if test_helper.is_file():
copy(test_helper, 'backup-tools/test_expansion_backup.py')
copy(pilot / 'best.pt', artifact_directory + '/best.pt', selected_hash)
copy(pilot / 'checkpoint.pt', artifact_directory + '/checkpoint.pt', checkpoint_hash)
prediction_name = artifact_directory + f"/validation_step_{best['step']:06d}_predictions.json"
copy(predictions_path, prediction_name)
for name in ('manifest.json', 'summary.json', 'correctness_initial.json', 'correctness_final.json',
'data_filter.json', 'best_validation_selection.json'):
copy(pilot / name, artifact_directory + '/' + name)
copy(pilot.parent / 'exit_code', artifact_directory + '/exit_code')
parent_directory = 'artifacts/warm-start-parent'
copy(parent_path, parent_directory + '/best.pt', parent_hash)
parent_prediction_name = parent_directory + '/' + parent_predictions.name
copy(parent_predictions, parent_prediction_name)
for name in ('manifest.json', 'correctness_initial.json', 'data_filter.json', 'best_validation_selection.json',
'parent-snapshot.json', 'source-data-manifest.json'):
copy(parent_path.parent / name, parent_directory + '/' + name)
data_prefix = 'data/' + dataset.name
copy(dataset / 'manifest.json', data_prefix + '/manifest.json', expected_transition['manifest_sha256'])
copy(base_root / 'manifest.json', 'data/' + base_root.name + '-manifest.json', expected_transition['parent_manifest_sha256'])
for split in SPLITS:
copy(dataset / (split + '.jsonl'), data_prefix + '/' + split + '.jsonl', data['split_sha256'][split])
copy(dataset / diagnostic_name, data_prefix + '/' + diagnostic_name, diagnostics['sha256'])
proof_directory = dataset.with_name(dataset.name + '-proof')
if proof_directory.is_dir():
for name in ('tokenization-proof.json', 'data_filter.json', 'postbuild-audit.json',
'prepare_expanded_data.py', 'tokenize_only.py'):
copy(proof_directory / name, data_prefix + '/proof/' + name)
for family, info in data['sources'].items():
if 'license_reference' in info:
recorded = info['files'][info['license_reference']]
relative = relative_name(recorded['local_path'])
copy(dataset / relative, data_prefix + '/' + relative, recorded['sha256'])
notices = ['# Dataset attribution and transformations', '',
'This private backup retains the upstream licences recorded in the dataset manifest.',
'No blanket licence is assigned to this mixed-source collection or to the model.',
'The transformed five splits and diagnostics are copied exactly; raw source datasets and token caches are omitted.',
'Source pins, raw-file hashes, URLs, normalization, sampling and grouping transformations are recorded in manifest.json.',
'Calibration/test/holdout files are copied as opaque bytes; their predictions are never used by this helper.', '',
'| Family | Upstream repository | Pinned revision | Recorded licence |', '| --- | --- | --- | --- |']
notices += [f"| {family} | {info['repo']} | {info['revision']} | {info['license']} |"
for family, info in sorted(data['sources'].items())]
notice_path = staging / data_prefix / 'ATTRIBUTION.md'
notice_path.write_text('\n'.join(notices) + '\n')
files[str(notice_path.relative_to(staging))] = {'bytes': notice_path.stat().st_size, 'sha256': digest(notice_path)}
artifact = {'name': 'expanded-gx10-4b-pilot', 'source_training_directory': str(pilot),
'artifact_directory': artifact_directory, 'source_commit': revision,
'source_file_sha256': manifest['source_sha256'], 'base_model': best['model_provenance'],
'data_signature': best['data_signature'], 'initialization': best['initialization'],
'selected': {'path': artifact_directory + '/best.pt', 'step': best['step'], 'sha256': selected_hash,
'predictions_path': prediction_name, 'raw_macro_nll': metrics['macro_nll'],
'selection_metric': metrics['selection_metric'], 'selection_score': metrics['selection_score'],
'accuracy': metrics['accuracy'], 'prediction_count': metrics['count']},
'resumable': {'path': artifact_directory + '/checkpoint.pt', 'step': latest['step'],
'sha256': checkpoint_hash, 'optimizer_steps': optimizer_steps,
'paired_selected_best_step': best['step'], 'config': latest['config']}}
parent_artifact = {'name': 'warm-start-parent', 'role': 'weights-only initialization; no optimizer carry-over',
'source_training_directory': str(parent_path.parent), 'artifact_directory': parent_directory,
'source_commit': parent_revision, 'source_file_sha256': parent_manifest['source_sha256'],
'base_model': parent['model_provenance'], 'data_signature': parent['data_signature'],
'selected': {'path': parent_directory + '/best.pt', 'step': parent['step'], 'sha256': parent_hash,
'predictions_path': parent_prediction_name, 'raw_macro_nll': parent_metrics['macro_nll'],
'selection_metric': parent_metrics['selection_metric'], 'selection_score': parent_metrics['selection_score'],
'accuracy': parent_metrics['accuracy'], 'prediction_count': parent_metrics['count']}}
backup = {'format': 'opensysone-backup-v1', 'created_utc': datetime.now(timezone.utc).isoformat(),
'target_repository': repo_id, 'staging_directory': str(output), 'uploaded': False,
'purpose': 'Completed expanded-training pilot with selected and resumable weights plus exact transformed dataset',
'calibrated': False, 'reserved_predictions_accessed': False, 'gpu_initialized_during_backup': False,
'dataset': {'path': data_prefix, 'payload_included': True, 'diagnostics_selection_eligible': False,
'manifest_sha256': expected_transition['manifest_sha256'], 'split_sha256': data['split_sha256']},
'validation_reference': {'path': 'fleet/reference-validation-predictions.json',
'sha256': plan['reference_sha256'], 'source_path': str(reference_path)},
'fleet_plan': {'path': 'fleet/plan.json', 'sha256': plan_hash, 'captured_utc': plan_captured_utc},
'artifacts': [artifact, parent_artifact], 'files': files, 'total_payload_bytes': sum(item['bytes'] for item in files.values()),
'helper_sha256': digest(__file__),
'restore_requirements': ['Restore the pinned base weights separately.',
'Use the archived exact training source revision and recreate dataset paths or verify signatures after path relocation.',
'Keep checkpoint.pt, best.pt and matching validation evidence together; checkpoint.pt includes Adam/RNG state.',
'Each upstream dataset retains its recorded licence; no common project or dataset licence is asserted.']}
write_json(staging / 'backup-manifest.json', backup)
sums = {name: record['sha256'] for name, record in files.items()}
sums['backup-manifest.json'] = digest(staging / 'backup-manifest.json')
(staging / 'SHA256SUMS').write_text(''.join(f'{checksum} {name}\n' for name, checksum in sorted(sums.items())))
verify_artifacts(staging, repo_id)
if torch.cuda.is_initialized():
raise ValueError('Backup unexpectedly initialized CUDA')
for path in staging.rglob('*'):
if path.is_file():
path.chmod(0o444)
staging.rename(output)
return backup
except BaseException:
shutil.rmtree(staging)
raise
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--pilot', required=True, help='Completed local pilot run or its artifacts directory')
parser.add_argument('--output', required=True, help='Fresh immutable artifact directory outside the source checkout')
parser.add_argument('--fleet-plan', required=True)
parser.add_argument('--repo-id', default='andyshu/opensysone')
args = parser.parse_args()
Path('/proc/self/oom_score_adj').write_text('0')
try:
result = prepare(args.pilot, args.output, args.fleet_plan, args.repo_id)
except Exception as error:
print(json.dumps({'status': 'failed', 'error_type': type(error).__name__}), file=sys.stderr)
return 1
print(json.dumps({'status': 'prepared', 'output': args.output, 'payload_bytes': result['total_payload_bytes'],
'source_commit': result['artifacts'][0]['source_commit'], 'uploaded': False}))
return 0
if __name__ == '__main__':
raise SystemExit(main())
|