File size: 6,865 Bytes
7123a00 | 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 | """Single entry point for independent crop and native-fullpage checkpoint evaluation."""
import argparse
import hashlib
import json
import os
from pathlib import Path
import subprocess
import sys
EXP = Path(__file__).resolve().parent
V5 = EXP.parent.parent
PAGE = Path('/autodl-tmp/official_table_pages_20261031')
PACK = PAGE / 'hf_sync/canonical_rebalanced'
CROP_PY = '/root/autodl-tmp/venv_train/bin/python'
PAGE_PY = '/root/autodl-tmp/venv_paddleocrvl_eval/bin/python'
def dump(path, obj):
path.parent.mkdir(parents=True, exist_ok=True)
tmp = path.with_suffix('.tmp')
tmp.write_text(json.dumps(obj, ensure_ascii=False, indent=2) + '\n')
tmp.replace(path)
def digest(path):
return hashlib.sha256(path.read_bytes()).hexdigest()
def plan(model, tag, track):
crop_env = dict(os.environ,
PYTHONPATH='/root/autodl-tmp/vllm_local/site-packages:/root/autodl-tmp/venv_train/lib/python3.12/site-packages:/autodl-fs/data/vllm_env/lib/python3.12/site-packages',
POVL_VLLM='/root/autodl-tmp/vllm_local/bin/vllm', POVL_PARALLEL_ENGINE_BOOT='1',
TOKENIZERS_PARALLELISM='false', OMP_NUM_THREADS='4', MKL_NUM_THREADS='4')
crop_env.pop('POVL_RUN_DIR', None)
page_env = dict(crop_env)
page_env.pop('PYTHONPATH', None)
jobs = []
if track in ('crop', 'both'):
jobs.append(('crop', [CROP_PY, str(V5 / 'src/eval_mix.py'), '--model', model,
'--tag', tag + '_crop', '--packs', str(PACK), '--gpus', '0', '1', '2', '3',
'--per-gpu', '3', '--conc-per-worker', '16', '--score-workers', '12',
'--score-batch', '64', '--max-tokens', '16384'], crop_env))
if track in ('fullpage', 'both'):
jobs.append(('fullpage', [PAGE_PY, str(EXP / 'whole_page_eval.py'), '--model', model,
'--tag', tag + '_fullpage'], page_env))
return jobs
def crop_complete(directory):
scores, summary = directory / 'scores.jsonl', directory / 'summary.json'
if not scores.exists() or not summary.exists():
return False
keys = set()
for line in scores.open():
r = json.loads(line)
key = r['cohort'], r['sample_id']
if key in keys:
raise RuntimeError('Duplicate crop score identity: ' + str(key))
keys.add(key)
return len(keys) == 176810
def check_free_gpus():
free = subprocess.check_output(['nvidia-smi', '--query-gpu=memory.free',
'--format=csv,noheader,nounits'], text=True).splitlines()
if len(free) != 4 or min(map(int, free)) < 40000:
raise RuntimeError('Crop evaluation needs four free GPUs with >=40000 MiB each; existing training is left running.')
def main():
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument('--model', required=True, help='Complete PaddleOCR-VL merged checkpoint directory')
ap.add_argument('--tag', required=True, help='Unique checkpoint/run name')
ap.add_argument('--track', choices=['both', 'crop', 'fullpage'], default='both')
ap.add_argument('--output-dir', type=Path, default=Path('/autodl-tmp/table_evaluation'))
ap.add_argument('--dry-run', action='store_true', help='Print commands only; no writes or GPU work')
args = ap.parse_args()
if not args.tag or Path(args.tag).name != args.tag or args.tag in ('.', '..'):
ap.error('--tag must be one directory name')
model_path = Path(args.model).resolve()
config = model_path / 'config.json'
if not config.exists() or json.loads(config.read_text()).get('model_type') != 'paddleocr_vl':
ap.error('--model must contain a PaddleOCR-VL config.json')
weights = sorted(model_path.glob('*.safetensors'))
if not weights or (model_path / 'adapter_config.json').exists():
ap.error('Merge the adapter with the base first; a complete model is required')
jobs = plan(str(model_path), args.tag, args.track)
if args.dry_run:
print(json.dumps({'model': str(model_path), 'track': args.track,
'gpus': 4, 'vllm_instances_per_gpu': 3,
'commands': {name: cmd for name, cmd, env in jobs},
'inherited_training_run_removed': all('POVL_RUN_DIR' not in env for _, _, env in jobs),
'fullpage_pythonpath_removed': all('PYTHONPATH' not in env for name, _, env in jobs if name == 'fullpage'),
'summary': str(args.output_dir / args.tag / 'results.json')}, indent=2))
return
root = args.output_dir / args.tag
identity = {'model': str(model_path), 'config_sha256': digest(config),
'weights': [{'name': p.name, 'bytes': p.stat().st_size, 'mtime_ns': p.stat().st_mtime_ns} for p in weights],
'crop_release_sha256': digest(PACK / 'release_audit.json')}
binding = root / 'identity.json'
if binding.exists() and json.loads(binding.read_text()) != identity:
raise RuntimeError('Model or evaluation pack changed; use a new --tag')
# An unbound old crop cache cannot safely be associated with this model.
crop_dir = V5 / 'eval/runs' / (args.tag + '_crop')
if not binding.exists() and args.track in ('crop', 'both') and crop_dir.exists():
raise RuntimeError('Existing crop tag has no unified model binding; use a new --tag')
dump(binding, identity)
report_path = root / 'results.json'
report = json.loads(report_path.read_text()) if report_path.exists() else {'model': str(model_path),
'tracks': {}, 'combined_metric': None, 'note': 'Crop/fullpage metrics remain independent.'}
for name, cmd, env in jobs:
dump(root / 'status.json', {'phase': name, 'command': cmd})
if name == 'crop' and crop_complete(crop_dir):
print('Reusing completed, identity-bound crop scores', flush=True)
else:
if name == 'crop':
check_free_gpus()
print('Evaluating ' + name, flush=True)
with (root / (name + '.log')).open('a') as log:
subprocess.run(cmd, cwd=V5, env=env, stdout=log, stderr=subprocess.STDOUT, check=True)
summary = crop_dir / 'summary.json' if name == 'crop' else PAGE / (args.tag + '_fullpage') / 'SUMMARY.json'
if name == 'crop' and not crop_complete(crop_dir):
raise RuntimeError('Crop evaluation incomplete; inspect ' + str(root / 'crop.log'))
result = json.loads(summary.read_text())
if name == 'fullpage' and not result.get('complete'):
raise RuntimeError('Fullpage evaluation incomplete; inspect ' + str(root / 'fullpage.log'))
report['tracks'][name] = {'summary_file': str(summary), 'summary': result}
dump(report_path, report)
dump(root / 'status.json', {'phase': 'complete', 'tracks': [name for name, _, _ in jobs],
'results': str(report_path)})
print(str(report_path), flush=True)
if __name__ == '__main__':
main()
|