Files
MeltMind/tools/summarize_s2_pilot_checkpoint.py
T

46 lines
2.9 KiB
Python

"""Publish the fixed first-50 evidence without changing the running queue."""
import argparse
import json
import math
from pathlib import Path
import numpy as np
from freeze_g1_components import sha
from run_s2_pilot_preparation import projection
def summarize(root):
c=json.loads((root/'contract.json').read_text());checkpoint=json.loads((root/'first-50-checkpoint.json').read_text())
assert checkpoint['completed']==50
jobs=[]
for i in range(50):
p=root/'jobs'/f'job-{i:04d}'/'guard-result.json';r=json.loads(p.read_text())
jobs.append(dict(index=i,case=c['cases'][i]['case'],receipt_sha256=sha(p),**r))
projected=projection(c,jobs)
assert math.isclose(projected['projected_worker_cpu_seconds'],checkpoint['projected_worker_cpu_seconds'],rel_tol=1e-12,abs_tol=1e-8)
nodes=np.array([s['tet_nodes'] for s in c['cases']]);cuts=np.quantile(nodes,[1/3,2/3])
def distribution(a):
return dict(parts=len(a),nodes=int(a.sum()),mean=float(a.mean()),
quantiles=dict(zip(['min','p10','median','p90','max'],np.quantile(a,[0,.1,.5,.9,1]).tolist())),
full_pool_tertile_counts=np.bincount(np.searchsorted(cuts,a),minlength=3).tolist())
hours=projected['projected_worker_cpu_seconds']/3600;disk=checkpoint['projected_disk_bytes']/2**30
if hours>80 or disk>100:decision='STOP_AND_REPORT_OVER_NEW_CAP'
elif hours>60:decision='ORIGINAL_QUEUE_STOPS_FREEZE_NEVER_ATTEMPTED_CONTINUATION_WITHIN_80H'
else:decision='ORIGINAL_QUEUE_CONTINUES_WITHIN_60H'
report=dict(status='FIRST_50_TERMINAL_REVIEW',dispatch='frozen ascending case-ID order; unchanged',
contract_sha256=sha(root/'contract.json'),checkpoint_sha256=sha(root/'first-50-checkpoint.json'),
fixed_first50=distribution(nodes[:50]),full_pool=distribution(nodes),first50_node_share=float(nodes[:50].sum()/nodes.sum()),
tertile_cut_nodes=cuts.tolist(),jobs=jobs,failed=sum(r['worker']['status']!='PASS' for r in jobs),
worker_cpu_seconds=sum(r['worker']['cpu_seconds'] for r in jobs),guard_cpu_seconds=sum(r['supervisor_cpu_seconds'] for r in jobs),
peak_individual_worker_tree_rss_bytes=max(r['worker']['peak_tree_rss_bytes'] for r in jobs),
projection=projected,projected_worker_cpu_hours=hours,projected_gib=disk,
original_supervisor_cap_cpu_hours=60,new_ruling_cap_cpu_hours=80,new_ruling='1794682712.505709',
unchanged_order_and_continuation_ruling='1794683049.057259',decision=decision,
running_contract_unchanged=True,retry=False,replacement=False,test_labels_read=False,GPU_training_started=False)
with (root/'public/first50-review.json').open('x') as f:json.dump(report,f,indent=2);f.write('\n')
print(json.dumps({k:v for k,v in report.items() if k!='jobs'}))
return report
if __name__=='__main__':
p=argparse.ArgumentParser();p.add_argument('root',type=Path);a=p.parse_args();summarize(a.root)