211 lines
12 KiB
Python
211 lines
12 KiB
Python
"""Single O1 geometry queue with first-ten gate and bounded, non-retry attempts."""
|
|
import argparse
|
|
import fcntl
|
|
import json
|
|
import math
|
|
import os
|
|
from pathlib import Path
|
|
import signal
|
|
import subprocess
|
|
import sys
|
|
import time
|
|
|
|
import psutil
|
|
|
|
from o1_partition import checked, sha, common_roster, can_reserve, InputViolation
|
|
from o1_prepare import contract, read
|
|
from r4bc_guard import supervise
|
|
from s2_recovery_supervisor import save, tree_usage
|
|
from s2_recovery_monitoring import disk_usage
|
|
|
|
|
|
def part_dir(root, c, index):
|
|
return Path(root)/'parts'/c['parts'][index]['case']
|
|
|
|
|
|
def guard(root, index, slot):
|
|
c = contract(root); os.sched_setaffinity(0, [c['worker_cpus'][slot]])
|
|
dest = part_dir(root, c, index)
|
|
receipt = supervise([sys.executable, 'tools/o1_prepare.py', str(root), str(index)], dest,
|
|
c['part_limits'], global_stop=root/'global-stop.json')
|
|
output = dest/'output/result.json'; failure = dest/'output/failure.json'
|
|
failed = read(failure) if failure.exists() else None
|
|
unexplained = receipt['status'] != 'PASS' and not receipt['stop_reason'] and failed is None
|
|
global_failure = bool(receipt['monitor_errors']) or bool(failed and failed['global_stop']) or unexplained
|
|
if receipt['status'] == 'PASS' and not output.exists():
|
|
receipt.update(status='FAILED', stop_reason='MISSING_RESULT'); global_failure = True
|
|
record = dict(index=index,case=c['parts'][index]['case'],worker=receipt,
|
|
result_path=str(output) if output.exists() else None,
|
|
result_sha256=sha(output) if output.exists() else None,
|
|
global_failure=global_failure,guard_cpu_seconds=time.process_time())
|
|
save(dest/'guard-result.json', record)
|
|
return 0 if receipt['status'] == 'PASS' else 1
|
|
|
|
|
|
def cost_check(c, records):
|
|
by = {r['index']:r for r in records}
|
|
samples = []; rows = []
|
|
for i in c['first_ten']:
|
|
r = by.get(i)
|
|
if r is None:
|
|
return dict(passed=False, reason='FIRST_TEN_INCOMPLETE')
|
|
result = read(checked(r['result_path'],r['result_sha256'])) if r['result_path'] else {}
|
|
rows.append(dict(case=r['case'],worker_status=r['worker']['status'],
|
|
part_status=result.get('status'),reason=result.get('reason'),
|
|
cpu_seconds=r['worker']['cpu_seconds'],tets=c['parts'][i]['tets'],
|
|
candidate_calls=result.get('candidate_calls',0)))
|
|
if r['worker']['status']=='PASS' and result.get('candidate_calls')==32:
|
|
samples.append(r['worker']['cpu_seconds']/c['parts'][i]['tets'])
|
|
spent = c['initial_preparation_cpu_seconds']+sum(r['worker']['cpu_seconds'] for r in records)
|
|
if not samples:
|
|
return dict(passed=False,reason='NO_COMPLETE_32_SEED_COST_SAMPLE',samples=rows,spent_cpu_seconds=spent)
|
|
rate = max(samples)
|
|
remaining = sum(c['parts'][i]['tets'] for i in c['dispatch'] if i not in by)*rate
|
|
return dict(passed=bool(spent+remaining <= c['limits']['preparation_cpu_seconds']),
|
|
reason='PROJECTED_WITHIN_CAP' if spent+remaining <= c['limits']['preparation_cpu_seconds'] else 'COST_PROJECTION_STOP',
|
|
samples=rows,measured_max_cpu_per_tet=rate,remaining_projected_cpu_seconds=remaining,
|
|
spent_cpu_seconds=spent,total_projected_cpu_seconds=spent+remaining,
|
|
method='Maximum measured complete-32-candidate CPU/tet times remaining tet counts; heuristic only; hard guards remain authoritative.')
|
|
|
|
|
|
def freeze_results(root, c, records, complete):
|
|
results = {p['case']:dict(status='INELIGIBLE',reason=p['metadata_failure']) for p in c['parts'] if p.get('metadata_failure')}
|
|
for r in records:
|
|
if r['worker']['status']=='PASS':
|
|
result = read(checked(r['result_path'],r['result_sha256']))
|
|
else:
|
|
result = dict(status='WORKER_FAILED',reason=r['worker']['stop_reason'])
|
|
results[r['case']] = result
|
|
for p in c['parts']:
|
|
results.setdefault(p['case'],dict(status='NEVER_ATTEMPTED'))
|
|
save(root/'all-part-results.json',results)
|
|
if not complete:
|
|
return None
|
|
kept, bad = common_roster(c['parts'],results)
|
|
groups = sorted({p['component_id'] for p in kept})
|
|
strengths = {}
|
|
for group in groups:
|
|
rows = [results[p['case']] for p in kept if p['component_id']==group]
|
|
strengths[group] = {key:sum(r['diagnostics'][key]['interface']['value'] for r in rows)/len(rows) for key in ['T/G','T/C']}
|
|
strong = sum(v['T/G']>=.1 and v['T/C']>=.1 for v in strengths.values())
|
|
output = dict(status='GEOMETRY_COMMON_ROSTER_FROZEN_NO_FITS',original_parts=617,original_components=209,
|
|
parts=kept,paired_components=groups,excluded_components=bad,
|
|
component_interface_strength=strengths,strong_components=strong,
|
|
coverage_gate=len(groups)>=168,intervention_gate=strong>=math.ceil(.7*len(groups)) if groups else False,
|
|
retained_parts=len(kept),retained_nodes=sum(p['nodes'] for p in kept),
|
|
all_part_results_sha256=sha(root/'all-part-results.json'),
|
|
results={r['case']:dict(path=r['result_path'],sha256=r['result_sha256']) for r in records},
|
|
fit_started=False,labels_read=False)
|
|
save(root/'common-roster.json',output)
|
|
return sha(root/'common-roster.json')
|
|
|
|
|
|
def run(root):
|
|
root = Path(root)
|
|
with (root/'queue.lock').open('a') as lock:
|
|
fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
|
|
return locked(root)
|
|
|
|
|
|
def locked(root):
|
|
c=contract(root); launch=read(root/'launch-freeze.json')
|
|
checked(root/'contract.json',launch['contract_sha256'])
|
|
if (root/'started.json').exists(): raise FileExistsError('O1 queue already attempted; no restart')
|
|
os.sched_setaffinity(0,[1]); start=time.monotonic()
|
|
save(root/'started.json',dict(pid=os.getpid(),unix=time.time(),contract_sha256=sha(root/'contract.json'),
|
|
implementation_commit=launch['implementation_commit'],cpu_affinity=sorted(os.sched_getaffinity(0)),
|
|
mode='GEOMETRY_PREPARATION_ONLY',fit_started=False))
|
|
active={};records=[];cursor=0;gate=False;cut=None;reason=None;errors=[];accounting=True;requested=[None]
|
|
def interrupt(sig,frame): requested[0]='COORDINATOR_SIGNAL_'+str(sig)
|
|
for s in (signal.SIGTERM,signal.SIGINT): signal.signal(s,interrupt)
|
|
env=dict(os.environ,OMP_NUM_THREADS='1',OPENBLAS_NUM_THREADS='1',MKL_NUM_THREADS='1',NUMEXPR_NUM_THREADS='1',CUDA_VISIBLE_DEVICES='')
|
|
try:
|
|
while True:
|
|
for slot,item in list(active.items()):
|
|
if item['process'].poll() is None: continue
|
|
path=part_dir(root,c,item['index'])/'guard-result.json'
|
|
if not path.exists():
|
|
accounting=False; reason='MISSING_GUARD_RECEIPT'
|
|
else:
|
|
r=read(path)
|
|
if not math.isfinite(r['worker']['cpu_seconds']) or r['worker']['cpu_seconds']<0:
|
|
accounting=False;reason='INVALID_ACCOUNTING'
|
|
records.append(r)
|
|
if r['global_failure']: reason='WORKER_GLOBAL_FAILURE'
|
|
item['log'].close();del active[slot]
|
|
spent=c['initial_preparation_cpu_seconds']+sum(r['worker']['cpu_seconds'] for r in records)
|
|
reason=reason or requested[0]
|
|
try:
|
|
usage=[tree_usage(item['process'].pid) for item in active.values()]
|
|
live=sum(u[0] for u in usage);rss=sum(u[1] for u in usage)+psutil.Process().memory_info().rss
|
|
disk=disk_usage(root);available=psutil.virtual_memory().available
|
|
if live+spent>c['limits']['preparation_cpu_seconds']: reason='GLOBAL_PREPARATION_CPU_LIMIT'
|
|
if rss>c['limits']['total_rss_bytes']: reason='GLOBAL_RSS_LIMIT'
|
|
if disk>=c['limits']['disk_bytes']: reason='GLOBAL_DISK_LIMIT'
|
|
except Exception as exc:
|
|
errors.append(dict(type=type(exc).__name__,message=str(exc)));reason='GLOBAL_MONITOR_ERROR'
|
|
live=rss=disk=available=0
|
|
if not gate and not active and len(records)==len(c['first_ten']) and not reason and not cut:
|
|
check=cost_check(c,records);save(root/'first-ten-cost-check.json',check)
|
|
gate=check['passed']
|
|
if not gate: cut=check['reason']
|
|
waiting=None
|
|
if not reason and not cut:
|
|
while cursor<len(c['dispatch']) and len(active)<(4 if gate else 1):
|
|
if not gate and cursor>=len(c['first_ten']):break
|
|
if not can_reserve(spent,len(active),c['limits']['preparation_cpu_seconds']):
|
|
if not active: cut='NOT_RUN_BUDGET'
|
|
else: waiting='WAIT_ACTIVE_BUDGET_SETTLEMENT'
|
|
break
|
|
if available<10*2**30: waiting='WAIT_AVAILABLE_MEMORY';break
|
|
contract(root)
|
|
index=c['dispatch'][cursor];slot=next(i for i in range(4) if i not in active)
|
|
dest=part_dir(root,c,index);dest.mkdir(parents=True,exist_ok=True)
|
|
if (dest/'worker.started.json').exists(): raise InputViolation('Attempt already exists')
|
|
log=(dest/'guard.log').open('x')
|
|
proc=subprocess.Popen([sys.executable,'tools/run_o1_prepare.py','guard',str(root),str(index),str(slot)],
|
|
stdout=log,stderr=subprocess.STDOUT,env=env,start_new_session=True)
|
|
active[slot]=dict(process=proc,index=index,log=log);cursor+=1;available-=8*2**30
|
|
if reason:
|
|
if not (root/'global-stop.json').exists():save(root/'global-stop.json',dict(reason=reason,unix=time.time()))
|
|
for item in active.values():
|
|
if not item.get('stop_sent'):
|
|
try:item['process'].send_signal(signal.SIGTERM)
|
|
except ProcessLookupError:pass
|
|
item['stop_sent']=True
|
|
save(root/'progress.json',dict(terminal_attempts=len(records),metadata_excluded=sum(bool(p.get('metadata_failure')) for p in c['parts']),
|
|
active={str(s):dict(index=v['index'],case=c['parts'][v['index']]['case'],guard_pid=v['process'].pid) for s,v in active.items()},
|
|
preparation_cpu_seconds=spent,live_guard_and_worker_cpu_seconds=live,reserved_active_cpu_seconds=len(active)*1800,
|
|
total_rss_bytes=rss,disk_bytes=disk,first_ten_gate_passed=gate,cut=cut,global_stop=reason,waiting=waiting,
|
|
monitor_errors=errors,wall_seconds=time.monotonic()-start,fit_started=False))
|
|
if not active and (reason or cut or cursor==len(c['dispatch'])):break
|
|
time.sleep(.2)
|
|
except Exception as exc:
|
|
reason='COORDINATOR_EXCEPTION'
|
|
save(root/'coordinator-failure.json',dict(type=type(exc).__name__,message=str(exc)))
|
|
save(root/'global-stop.json',dict(reason=reason))
|
|
finally:
|
|
for item in active.values():
|
|
try:item['process'].send_signal(signal.SIGTERM)
|
|
except ProcessLookupError:pass
|
|
for item in active.values():
|
|
item['process'].wait();item['log'].close();p=part_dir(root,c,item['index'])/'guard-result.json'
|
|
if p.exists():records.append(read(p))
|
|
else:accounting=False
|
|
complete=not reason and not cut and accounting and len(records)==len(c['dispatch'])
|
|
roster_hash=freeze_results(root,c,records,complete)
|
|
save(root/'complete.json',dict(status='PREPARATION_COMPLETE' if complete else 'PREPARATION_STOPPED',
|
|
global_stop=reason,cut=cut,accounting_complete=accounting,records=records,
|
|
never_attempted=[c['parts'][i]['case'] for i in c['dispatch'][cursor:]],
|
|
preparation_cpu_seconds=c['initial_preparation_cpu_seconds']+sum(r['worker']['cpu_seconds'] for r in records),
|
|
guard_cpu_seconds=sum(r.get('guard_cpu_seconds',0) for r in records),coordinator_cpu_seconds=time.process_time(),
|
|
common_roster_sha256=roster_hash,monitor_errors=errors,fit_started=False,new_label_decodes=0,
|
|
next='No auto fit launch; R4bc terminal dependency and model implementation still required'))
|
|
return 0 if complete else 1
|
|
|
|
|
|
if __name__=='__main__':
|
|
p=argparse.ArgumentParser();p.add_argument('mode',choices=['run','guard']);p.add_argument('root',type=Path)
|
|
p.add_argument('index',type=int,nargs='?');p.add_argument('slot',type=int,nargs='?');a=p.parse_args()
|
|
raise SystemExit(run(a.root) if a.mode=='run' else guard(a.root,a.index,a.slot))
|