Files
MeltMind/tools/phase3_data_sufficiency_v2_guard.py
T

87 lines
5.7 KiB
Python

"""Independent bounded ITEM1 fold guard; shared global stop and immutable S1."""
import fcntl
import json
import os
from pathlib import Path
import signal
import sys
import threading
import time
import psutil
from phase3_data_sufficiency_v2_common import config
from phase2_h1_worker_v2 import save,sha
from s2_main_memory_guard import PeakRecords
from s2_precision_continuation_guard import device_usage
from s2_recovery_gpu_supervisor import supervise,tree_usage,stop_group,disk_usage
def run(queue,fold_id):
c,_=config(queue);job=c['jobs'][fold_id];c=dict(c,**job);root=queue/'jobs'/str(fold_id);os.sched_setaffinity(0,c['guard_cpus'])
with (root/'fold.lock').open('a') as lock:
fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
if (root/'started.json').exists():raise FileExistsError('Single ITEM1 fold launch, no retry')
start=time.monotonic();cpu=time.process_time();requested=[];reasons=[];errors=[];samples=[];done=threading.Event();tracker=PeakRecords();receipt=None
save(root/'started.json',dict(pid=os.getpid(),unix=time.time(),contract_sha256=sha(queue/'contract.json'),gpu_uuid=c['gpu_uuid'],setup_gpu_seconds=c['setup_gpu_seconds']))
def elapsed():return c['setup_gpu_seconds']+time.monotonic()-start
def reason(x):
if x not in reasons:reasons.append(x)
command=[sys.executable,'tools/phase3_data_sufficiency_v2_worker.py',str(queue),str(fold_id)]
def kill_worker():
for p in psutil.Process().children(recursive=True):
try:
if p.cmdline()==command:stop_group(p.pid)
except psutil.NoSuchProcess:pass
old={s:signal.signal(s,lambda signum,frame:requested.append('COORDINATOR_SIGNAL_'+str(signum))) for s in [signal.SIGINT,signal.SIGTERM]}
def scan():
if tracker.scan(root/'worker'):reason('RECORDED_VRAM_LIMIT')
def watch():
while not done.is_set():
try:
scan();mem,util=device_usage(c['gpu_uuid']);_,rss,_=tree_usage(os.getpid())
if mem>c['vram_bytes']:reason('DEVICE_VRAM_LIMIT')
if rss>c['rss_bytes']:reason('WHOLE_TREE_RSS_LIMIT')
if disk_usage(root)>c['disk_bytes']:reason('OUTPUT_DISK_LIMIT')
if elapsed()>=c['gpu_seconds']-5:reason('ITEM1_FOLD_TIME_LIMIT')
if (queue/'global-stop.json').exists():reason('ITEM1_GLOBAL_STOP')
for x in requested:reason(x)
save(root/'live-progress.json',dict(stage='ITEM1_TRAINING',fold=fold_id,elapsed_gpu_seconds=elapsed(),device_bytes=mem,device_utilization=util,tree_rss_bytes=rss,recorded_peaks=tracker.maxima,reasons=reasons,unix=time.time()))
except Exception as exc:reason('MONITOR_ERROR');errors.append(dict(type=type(exc).__name__,reason=str(exc)))
if reasons:kill_worker()
done.wait(.5)
thread=None
try:
for i in range(3):
mem,util=device_usage(c['gpu_uuid']);samples.append(dict(unix=time.time(),memory_bytes=mem,utilization=util))
if mem>128*2**20 or util!=0:raise RuntimeError('SELECTED_GPU_BUSY')
if i<2:time.sleep(2)
save(root/'idle-polls.json',samples)
thread=threading.Thread(target=watch,daemon=True);thread.start()
remaining=c['gpu_seconds']-elapsed()-10
receipt=supervise(command,root/'guards','train',dict(total_cpu_seconds=c['cpu_seconds']-60,job_cpu_seconds=c['cpu_seconds']-60,job_wall_seconds=remaining,tree_rss_bytes=c['rss_bytes'],disk_bytes=c['disk_bytes'],max_processes=1),poll=.2,gpu_uuid=c['gpu_uuid'])
except BaseException as exc:reason(type(exc).__name__+': '+str(exc));kill_worker()
finally:
done.set()
if thread:thread.join(timeout=4)
if thread and thread.is_alive():reason('MONITOR_DID_NOT_EXIT')
try:scan()
except Exception as exc:reason('FINAL_PEAK_SCAN_ERROR');errors.append(dict(reason=str(exc)))
for s,handler in old.items():signal.signal(s,handler)
value=None;p=root/'worker/result.json'
try:
if p.exists():value=json.loads(p.read_text())
except Exception as exc:reason('RESULT_READ_ERROR');errors.append(dict(reason=str(exc)))
if not receipt or receipt['status']!='PASS':reason('WORKER_NOT_PASS')
if receipt and receipt.get('monitor_errors'):reason('WORKER_MONITOR_ERRORS')
if not value or value.get('status')!='COMPLETE':reason('NO_COMPLETE_TRAIN_RESULT')
if not all(k in tracker.witnesses for k in ['peak_allocated_bytes','peak_reserved_bytes']):reason('MISSING_CUDA_PEAK_RECORDS')
totalcpu=60+(receipt['cpu_seconds'] if receipt else 0)+time.process_time()-cpu
if elapsed()>c['gpu_seconds'] or totalcpu>c['cpu_seconds']:reason('FINAL_RESOURCE_LIMIT')
try:config(queue)
except Exception as exc:reason('FINAL_SOURCE_IDENTITY: '+str(exc))
if value and (value.get('optimizer_steps')!=c['target_steps'] or value.get('job')!=fold_id):reason('INCOMPLETE_EPOCH_OR_FOLD_COUNT')
save(root/'complete.json',dict(status='PASS' if not reasons else 'STOPPED_OR_FAILED',decision='TRAIN_COMPLETE_CHECKPOINT_FREEZE_PENDING' if not reasons else 'ITEM1_FOLD_INCOMPLETE',reasons=reasons,monitor_errors=errors,
worker=receipt,recorded_memory=tracker.as_dict(),gpu_seconds_charged=elapsed(),fold=fold_id,cpu_seconds_charged=totalcpu,
worker_result_sha256=sha(p) if p.exists() else None,prediction_started=False,scoring_started=False,s2_started=True,no_retry=True))
return 0 if not reasons else 1
if __name__=='__main__':raise SystemExit(run(Path(sys.argv[1]),int(sys.argv[2])))