Files
MeltMind/tools/phase3_seed1_v2_observation.py
T

77 lines
4.8 KiB
Python

"""Epoch-boundary host observations only; no scheduling or causal decisions."""
import csv
import io
import json
import os
from pathlib import Path
import subprocess
import time
import psutil
from phase2_h1_worker_v2 import save,sha
def epoch_host_observation(root,fold,epoch,gpu_uuid,receipt=None,incomplete_reason=None):
record=dict(unix=time.time(),fold=fold,completed_epoch=epoch,assigned_gpu_uuid=gpu_uuid,observation_only=True,sampling='After the coordinator reads the completed-epoch receipt; not continuous monitoring and not causal proof',gpu_processes=[],errors=[])
if receipt is not None:
receipt=Path(receipt)
record.update(receipt_path=str(receipt),receipt_sha256=sha(receipt),receipt_mtime_unix=receipt.stat().st_mtime,receipt_age_seconds=max(0,record['unix']-receipt.stat().st_mtime))
if incomplete_reason is not None:
record.update(status='OBSERVATION_INCOMPLETE',incomplete_reason=incomplete_reason,snapshot_started_unix=None,snapshot_finished_unix=None,own_worker_pids=[],historical_gpu_occupancy_observed=False)
return write_observation(root,fold,epoch,record)
record['snapshot_started_unix']=time.time()
try:record['load_average']=list(os.getloadavg())
except OSError as exc:record['errors'].append(dict(stage='load_average',type=type(exc).__name__,reason=str(exc)))
own=set()
try:
for p in psutil.process_iter(['pid','cmdline']):
args=p.info['cmdline'] or []
if str(root) in args and any(Path(a).name=='phase3_seed1_v2_worker.py' for a in args):own.add(p.pid)
except Exception as exc:record['errors'].append(dict(stage='own_worker_identity',type=type(exc).__name__,reason=str(exc)))
try:
result=subprocess.run(['nvidia-smi','--query-compute-apps=gpu_uuid,pid,process_name,used_memory','--format=csv,noheader,nounits'],capture_output=True,text=True,timeout=5,check=True)
for row in csv.reader(io.StringIO(result.stdout)):
if not row:continue
if len(row)!=4:raise ValueError('Unexpected GPU process columns')
uuid,pid,name,memory=[s.strip() for s in row];pid=int(pid)
record['gpu_processes'].append(dict(gpu_uuid=uuid,pid=pid,process_name=name,used_memory_mib=memory,own_v2_worker=pid in own,on_assigned_gpu=uuid==gpu_uuid,foreign_on_assigned_gpu=uuid==gpu_uuid and pid not in own))
except Exception as exc:record['errors'].append(dict(stage='gpu_process_snapshot',type=type(exc).__name__,reason=str(exc)))
if not any(p['own_v2_worker'] and p['on_assigned_gpu'] for p in record['gpu_processes']):
record['errors'].append(dict(stage='epoch_worker_visibility',reason='Assigned v2 worker not visible at snapshot; historical occupancy not observed'))
record['own_worker_pids']=sorted(own);record['status']='OBSERVED' if not record['errors'] else 'OBSERVATION_INCOMPLETE'
record['snapshot_finished_unix']=time.time()
record['historical_gpu_occupancy_observed']=False
return write_observation(root,fold,epoch,record)
def write_observation(root,fold,epoch,record):
p=root/'jobs'/str(fold)/'epoch-host-observations'/f'{epoch:03d}.json'
p.parent.mkdir(parents=True,exist_ok=True)
if p.exists():raise FileExistsError('Preserve original epoch observation')
save(p,record)
return record
def drain_epoch_observations(root,fold,gpu_uuid,covered,terminal=False,stopping=False):
"""Account for each completed receipt independently of the cost cursor.
Terminal, stopping, batched or >5s-old receipts get explicit missing-history
records without GPU queries. A fresh live singleton gets a current snapshot;
its true times never claim coverage of earlier epoch occupancy.
"""
history=root/'jobs'/str(fold)/'worker/history'
receipts=[]
for p in sorted(history.glob('*.json')):
row=json.loads(p.read_text());epoch=row['epoch']
if type(epoch) is not int or not 1<=epoch<=200 or p.stem!=f'{epoch:03d}':raise ValueError('Completed epoch receipt identity')
if epoch not in covered:receipts.append((epoch,p))
backlog=len(receipts)>1;new_records=[]
for epoch,p in receipts:
output=root/'jobs'/str(fold)/'epoch-host-observations'/f'{epoch:03d}.json'
if output.exists():
record=json.loads(output.read_text())
if record['completed_epoch']!=epoch or record['fold']!=fold or record.get('receipt_sha256')!=sha(p):raise ValueError('Existing observation identity')
else:
reason=('TERMINAL_BEFORE_OBSERVATION' if terminal else 'GLOBAL_STOP_BEFORE_OBSERVATION' if stopping else 'MULTIPLE_COMPLETED_RECEIPTS_BACKLOG' if backlog else 'DELAYED_RECEIPT_OVER_5_SECONDS' if time.time()-p.stat().st_mtime>5 else None)
record=epoch_host_observation(root,fold,epoch,gpu_uuid,receipt=p,incomplete_reason=reason)
covered.add(epoch);new_records.append(record)
return new_records