Freeze hash-bound zero-run v2 continuation with recorded idle settling
This commit is contained in:
1 parent
a15f2882ee
commit
ea5ac05191
8 files changed
+366
No files matched your search
@@ -0,0 +1,79 @@
|
||||
{
|
||||
"version": "S2-v2-zero-run-engineering-continuation",
|
||||
"authorizing_ruling": "1794695421.177269",
|
||||
"source": "artifacts/s2-pilot-v2-2026-11-15",
|
||||
"source_commit": "a15f2882ee2e8c3392c6b7c80aa521d7ba2ad58e",
|
||||
"source_sha256": {
|
||||
"artifacts/s2-pilot-v2-2026-11-15/pilot-plan.json": "5a127d5ddfab366da49475867c9f14f2854f5000a733c7a3211df61bd7acecae",
|
||||
"artifacts/s2-pilot-v2-2026-11-15/training-contract.json": "0218b74f8d86ae4190e0b814f7c5c8a7224a616b94ad7587120156e755de2101",
|
||||
"artifacts/s2-pilot-v2-2026-11-15/timing-probe/result.json": "96b95ec8dd0e37c38096dda9ec06257eb02662f8fd62cbe11b9ad58887d8376b",
|
||||
"artifacts/s2-pilot-v2-2026-11-15/timing-probe-guard-receipt.json": "45e9fe236d31dd7eabbbd46edce4fc7a80914c5a12088d1977de3633d79d6c2a",
|
||||
"artifacts/s2-pilot-v2-2026-11-15/pilot-complete.json": "337ea9552569db135e22b93f3e0b1557e03c5b56669cdb4d1bbcd391e6f67982",
|
||||
"artifacts/s2-pilot-v2-2026-11-15/orchestrator-complete.json": "e2794244d0664cc10aa3707a8da4f2d078946f985a1095f82a2f12b5310128a9"
|
||||
},
|
||||
"code_sha256": {
|
||||
"tools/prepare_s2_recovered_pilot_training.py": "ab5a4274363598ff933ff873df2b00497f069b1fb45756a1d1dba8d723294f68",
|
||||
"tools/run_s2_recovered_pilot.py": "56962009bf5eaaee599ffea6a3d454372e0aea2842fa79eac609fd4a24ecbb39",
|
||||
"tools/run_s2_pilot_probe.py": "2360ab7b44b179144a5ecd2d691e33cd36a6ec4bfeac71452bf062e7f42f81eb",
|
||||
"tools/run_s2_pilot_training.py": "d5e543478543243fe1db3475a4d4cd135fbdcf631d3bce48c4b5af8b2edafb22",
|
||||
"tools/s2_pilot_train.py": "e9cceead51a5cc092aad23409ad98909c84a62e7e10571e750d1e9be1ad3b948",
|
||||
"tools/s2_recovered_pilot_gpu_guard.py": "9a27f47918c865ce37e46b0be54144dc3f9ab1058e4eddc9f6fbc771f4959eb0",
|
||||
"tools/s2_recovery_gpu_supervisor.py": "689c0f3179e4fcd362e546d90214829a9e4f8739abdbdf74b560763e5a392f15",
|
||||
"tools/s2_e1_supervisor.py": "6753c9ceb0b932047ba505bd15681f109e9407bdb58fd96c55e1431a3f1b4a02",
|
||||
"tools/s2_pilot_data.py": "82c1fd18ec51757479c8fd7dda2db4d5f38c366d77fedcca3216a08998214c97",
|
||||
"tools/s2_incident_recovery_manifest.py": "8cbb2e1067c2cc22762d8ae3cf192aa2c5fa7c7a199a22e7a39a25570ee1c3b1",
|
||||
"tools/freeze_g1_components.py": "d0cee60d95243ac2c59ef3f2244ba818315718c4e3d1895f2f9f064f1f7f5ecd",
|
||||
"tools/s2_recovery_monitoring.py": "442c2a9c75e8c00b30a8f2e91cb52d87c466b19a80aea9128b10c021e61b922e",
|
||||
"tools/prepare_s2_incident_recovery.py": "4f772573dedc2b7921986e25eaedb14f91d98b235421829c284f3b40b85710aa",
|
||||
"tools/run_s2_recovery_chain.py": "76d0808bdb2d624cfbaadf5a3c5e7e9a0c2be5d5ac41d4ac3fd9ccd48743933c",
|
||||
"tools/s2_models/__init__.py": "a0fa1323d3553172da66c5d442c5fc4fcb0d5055c886c776928604550ea97609",
|
||||
"tools/s2_models/attention.py": "7714adfccf6a0fc6b52a44aee3a50321a88898bf3d743680a1108c01e1ff8a61",
|
||||
"tools/s2_models/blocks.py": "b13e986cea5f1d7f325a1638c1505197b4a105baacff5a9a515f23fcaf28ea75",
|
||||
"tools/s2_models/features.py": "3b34ae61cf8012ebd58b7505a73d7aa43d8ee1890ac54f98d7917b67887a45a4",
|
||||
"tools/s2_models/network.py": "5137e845d81d1223afc13bf87e03ccba1132a3baa75d5189a8f320b2752be071",
|
||||
"tools/s2_models/normalization.py": "7180696d3845799e28d8c31ddef957d075de6a2e46d96a3a0f0df16fc05234a0",
|
||||
"tools/s2_models/objectives.py": "5b6527a45f08690055d572335470ce5979e8bb86e1cab00de594f79ae21316b1",
|
||||
"tools/prepare_s2_precision_pilot.py": "c3a71b0ff58bcea4433ff93e18ef14d3a00521438a76baaa69830a5ee8ac098a",
|
||||
"tools/run_s2_precision_pilot.py": "ac8347be1ca11f48cb67abbf5e5c78b39af46ed3a7c6f89b49098fd5e77a6904",
|
||||
"tools/s2_precision_pilot_gpu_guard.py": "5cf0f41de34b28394dbee4672b5bd34eb9c4fe2344822aa3d648f20561d96328",
|
||||
"tools/run_s2_pilot_probe_v2.py": "61c08499346d50405d0c254de54deea24b3522256bf1c8b0aec3e9f4ddab6390",
|
||||
"tools/run_s2_pilot_training_v2.py": "f461cb43bc602a32677d7dcfb8a322e206531ad152fa33bb4e36ea8546a1c0f2",
|
||||
"tools/run_s2_encoder_diagnostic.py": "8f31d312fb53bfdb9b1804af064e463d3989f1b33c4aa7279a967ca82dd2c0a4",
|
||||
"tools/s2_models_v2/__init__.py": "0302d333bcd8b33e85aaa6a4e7822e7e1ccfc66343136107fea6f4ca283ba1ec",
|
||||
"tools/s2_models_v2/attention.py": "3829e34ca30942c863699c4d43fb6fa104452c6c04c4d606c11019d3dde16ffa",
|
||||
"tools/s2_models_v2/blocks.py": "b13e986cea5f1d7f325a1638c1505197b4a105baacff5a9a515f23fcaf28ea75",
|
||||
"tools/s2_models_v2/network.py": "a443470caeecdf529504bffab3ca8cc0b70ce44b505bad60f9c63db9709dbb83",
|
||||
"tools/run_s2_precision_continuation.py": "9fb5a22007788f0c18c77a82b6a8b76fbb53a98b49cc9251bc49a7a5b8312f2f",
|
||||
"tools/s2_precision_continuation_guard.py": "01235d897c7c03e5ffa109c58b14f5d24d128e8f1ba873e1f76cfb14e02f7a7f",
|
||||
"tools/s2_gpu_idle_gate.py": "e2b1613e7e6ea398a8d555220a9365a1ad998ab4ef4b2bf8ec7cc4e31f8b4b9f"
|
||||
},
|
||||
"prior_wall_seconds": 11.143491394002922,
|
||||
"remaining_wall_seconds_at_entry": 7188.856508605997,
|
||||
"total_v2_wall_limit_seconds": 7200,
|
||||
"probe_reused": true,
|
||||
"probe_rerun": false,
|
||||
"E_max": 42,
|
||||
"cosine_T_max": 42,
|
||||
"training_runs_started_before_continuation": 0,
|
||||
"idle_gate": {
|
||||
"interval_seconds": 10,
|
||||
"consecutive_idle_required": 3,
|
||||
"selected_wait_limit_seconds": 600,
|
||||
"memory_MiB_at_most": 128,
|
||||
"utilization_percent": 0,
|
||||
"compute_processes_required": 0,
|
||||
"foreign_processes_never_killed": true,
|
||||
"record_PID_user_memory_every_poll": true,
|
||||
"switch_only_after_foreign_occupant_timeout_to_identical_V100": true
|
||||
},
|
||||
"all_new_waits_and_setup_in_original_7200_seconds": true,
|
||||
"original_busy_trigger_PID_user_memory": "UNKNOWN_NOT_RECORDED_AT_TRIGGER",
|
||||
"original_post_stop_observation": {
|
||||
"memory_MiB": 5,
|
||||
"utilization_percent": 0,
|
||||
"compute_processes": []
|
||||
},
|
||||
"scientific_code_changed": false,
|
||||
"test_labels_read": false,
|
||||
"no_automatic_retry": true
|
||||
}
|
||||
@@ -0,0 +1,6 @@
|
||||
{
|
||||
"status": "PASS",
|
||||
"tests_run": 7,
|
||||
"CPU_only": true,
|
||||
"observed_unix": 1794695603.3577545
|
||||
}
|
||||
@@ -0,0 +1,12 @@
|
||||
test_busy_then_three_consecutive_free (test_s2_gpu_idle_gate.IdleGateTests) ... ok
|
||||
test_foreign_occupant_times_out_without_killing (test_s2_gpu_idle_gate.IdleGateTests) ... ok
|
||||
test_short_remaining_budget_cannot_start_after_two_idle_polls (test_s2_gpu_idle_gate.IdleGateTests) ... ok
|
||||
test_exact_device_memory_units (test_s2_pilot_gpu_guard.GPUUsageTests) ... ok
|
||||
test_explicit_visibility_reaches_worker_and_receipt_without_gpu_use (test_s2_pilot_gpu_guard.GPUUsageTests) ... ok
|
||||
test_missing_or_ambiguous_counters_fail_closed (test_s2_pilot_gpu_guard.GPUUsageTests) ... ok
|
||||
test_vram_monitor_error_and_whole_host_rss_stop_without_second_run (test_s2_pilot_gpu_guard.GPUUsageTests) ... ok
|
||||
|
||||
----------------------------------------------------------------------
|
||||
Ran 7 tests in 0.036s
|
||||
|
||||
OK
|
||||
@@ -0,0 +1,11 @@
|
||||
# S2 v2 零训练run基础设施停止后的接续
|
||||
|
||||
Claude1794695421.177269接受v2 probe并裁量复用,不再probe。原session26792的PILOT_INCOMPLETE/runs=[]/SELECTED_GPU_BUSY_BEFORE_RUN及原11.143491394秒总账保持不变,固定提交a15f288。触发时PID、用户和占用值未由原门控保存,不能追溯确认占用者;停止后观测5MiB/0%/无compute进程不等于触发时证据。
|
||||
|
||||
新独立入口run_s2_precision_continuation.py绑定v2 plan、probe结果及guard、训练契约和两份终态SHA。根artifacts/s2-pilot-v2-continuation-2026-11-15。仅四训练run从头,不重新准备、probe或改变科学代码;E_max=cosine_T_max=42。原11.143491394秒已含probe,不重复扣probe;新入口剩余7188.856508606秒,全部新门控等待、设置和训练都计入同一7200秒。
|
||||
|
||||
入口对指定UUID每10秒采样,最多600秒,须连续3次memory≤128MiB、util0%、compute进程为空。每次保存PID/用户名/显存、设备值或采样错误;不杀其他进程。600秒后仍有外部占用才可查另一张名称完全相同的V100,并同样连续确认空闲,切换写入新的训练契约;否则停止。后续各run之间也使用同一门控,等待后重算剩余预算,不借新入口重置限额。
|
||||
|
||||
7项CPU合成检查覆盖busy→free、连续空闲被busy打断、外部占用600秒超时且不杀进程、余时不足不能通过3次检查,以及新guard的VRAM/主机内存/监控失败不进入第二run与可见性/计数器解析。原probe/precision源码pin不改。
|
||||
|
||||
实际启动和最终状态以新根started、idle-gates逐次JSONL、training-contract、guards/<run>.started、run/started及complete/failure为准。pilot-started只是入口标记,不代表训练worker已启动。
|
||||
@@ -0,0 +1,57 @@
|
||||
"""Engineering continuation of zero-run v2; reuse probe, preserve both stops."""
|
||||
import argparse
|
||||
import fcntl
|
||||
import json
|
||||
from pathlib import Path
|
||||
import time
|
||||
import traceback
|
||||
from freeze_g1_components import sha
|
||||
from s2_e1_supervisor import save
|
||||
from s2_gpu_idle_gate import inventory, recorded_wait
|
||||
from s2_precision_continuation_guard import run as train_runs
|
||||
|
||||
|
||||
def run(root):
|
||||
with (root/'continuation.lock').open('a') as lock:
|
||||
fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
|
||||
assert not (root/'started.json').exists(),'No automatic continuation retry'
|
||||
contract=json.loads((root/'continuation-contract.json').read_text())
|
||||
for p,h in contract['code_sha256'].items():assert sha(p)==h,p
|
||||
for p,h in contract['source_sha256'].items():assert sha(p)==h,p
|
||||
source=Path(contract['source']);old=json.loads((source/'training-contract.json').read_text())
|
||||
ended=json.loads((source/'pilot-complete.json').read_text())
|
||||
assert ended['status']=='PILOT_INCOMPLETE' and ended['reason']=='SELECTED_GPU_BUSY_BEFORE_RUN' and ended['runs']==[]
|
||||
previous=json.loads((source/'orchestrator-complete.json').read_text())['total_wall_seconds']
|
||||
assert previous==contract['prior_wall_seconds'] and old['E_max']==old['cosine_T_max']==42
|
||||
assert sha(old['timing_probe_receipt'])==old['timing_probe_sha256']
|
||||
started=time.monotonic();save(root/'started.json',dict(started_unix=time.time(),contract_sha256=sha(root/'continuation-contract.json'),prior_wall_seconds=previous))
|
||||
try:
|
||||
names=inventory();selected=old['gpu_uuid'];assert 'V100' in names[selected]
|
||||
gate=recorded_wait(root,'initial-selected',selected,min(600,7200-previous-(time.monotonic()-started)-5))
|
||||
switched=False
|
||||
if gate['status']!='IDLE_CONFIRMED':
|
||||
if not gate['last'].get('occupants'):raise RuntimeError('Selected GPU failed idle gate without identified foreign occupant; no switch')
|
||||
for uuid,name in sorted(names.items()):
|
||||
if uuid==selected or name!=names[selected]:continue
|
||||
remaining=7200-previous-(time.monotonic()-started)-5
|
||||
if remaining<=20:break
|
||||
alternative=recorded_wait(root,'alternative-'+uuid,uuid,min(600,remaining))
|
||||
if alternative['status']=='IDLE_CONFIRMED':selected=uuid;switched=True;break
|
||||
if not switched:raise RuntimeError('No identical idle V100 after foreign-occupant timeout')
|
||||
c=dict(old);c['gpu_uuid']=selected;c['code_sha256']=dict(old['code_sha256'],**contract['code_sha256'])
|
||||
c['engineering_continuation']=dict(source=str(source),source_sha256=contract['source_sha256'],
|
||||
prior_wall_seconds=previous,probe_reused=True,probe_rerun=False,gpu_switched=switched,
|
||||
original_gpu_uuid=old['gpu_uuid'],selected_gpu_uuid=selected,ruling='1794695421.177269')
|
||||
save(root/'training-contract.json',c)
|
||||
train_runs(root,prior_wall_seconds=previous+time.monotonic()-started)
|
||||
result=json.loads((root/'pilot-complete.json').read_text())
|
||||
save(root/'complete.json',dict(status=result['status'],selected_global_width=result['selected_global_width'],
|
||||
cumulative_v2_wall_seconds=previous+time.monotonic()-started,prior_wall_seconds=previous,
|
||||
limit_seconds=7200,probe_reused=True,probe_rerun=False,test_labels_read=False))
|
||||
except Exception as exc:
|
||||
save(root/'failure.json',dict(status='CONTINUATION_STOPPED_NO_RETRY',error=str(exc),traceback=traceback.format_exc(),
|
||||
cumulative_v2_wall_seconds=previous+time.monotonic()-started,prior_wall_seconds=previous,limit_seconds=7200));raise
|
||||
|
||||
|
||||
if __name__=='__main__':
|
||||
p=argparse.ArgumentParser();p.add_argument('root',type=Path);a=p.parse_args();run(a.root)
|
||||
@@ -0,0 +1,53 @@
|
||||
"""Bounded, recorded idle observations; never signal foreign processes."""
|
||||
import csv
|
||||
import json
|
||||
import subprocess
|
||||
import time
|
||||
from pathlib import Path
|
||||
import psutil
|
||||
|
||||
|
||||
def inventory():
|
||||
r=subprocess.run(['nvidia-smi','--query-gpu=uuid,name','--format=csv,noheader'],check=True,capture_output=True,text=True,timeout=3)
|
||||
return {row[0].strip():row[1].strip() for row in csv.reader(r.stdout.splitlines())}
|
||||
|
||||
|
||||
def snapshot(uuid):
|
||||
r=subprocess.run(['nvidia-smi','--id='+uuid,'--query-gpu=memory.used,utilization.gpu','--format=csv,noheader,nounits'],check=True,capture_output=True,text=True,timeout=3)
|
||||
rows=list(csv.reader(r.stdout.splitlines()));assert len(rows)==1
|
||||
memory,util=[int(v.strip()) for v in rows[0]];assert memory>=0 and 0<=util<=100
|
||||
r=subprocess.run(['nvidia-smi','--query-compute-apps=gpu_uuid,pid,process_name,used_memory','--format=csv,noheader,nounits'],check=True,capture_output=True,text=True,timeout=3)
|
||||
occupants=[]
|
||||
for row in csv.reader(r.stdout.splitlines()):
|
||||
if not row or row[0].strip()!=uuid:continue
|
||||
pid=int(row[1]);user=None
|
||||
try:user=psutil.Process(pid).username()
|
||||
except (psutil.NoSuchProcess,psutil.AccessDenied):pass
|
||||
occupants.append(dict(pid=pid,user=user,process_name=row[2].strip(),memory_MiB=row[3].strip()))
|
||||
return dict(unix=time.time(),uuid=uuid,memory_MiB=memory,utilization_percent=util,occupants=occupants,
|
||||
idle=memory<=128 and util==0 and not occupants)
|
||||
|
||||
|
||||
def wait_idle(uuid,limit_seconds,record,*,sample=snapshot,clock=time.monotonic,sleep=time.sleep):
|
||||
begin=clock();consecutive=0;count=0;last=None
|
||||
while True:
|
||||
try:last=sample(uuid)
|
||||
except Exception as exc:
|
||||
last=dict(uuid=uuid,unix=time.time(),idle=False,error=str(exc),occupants=[])
|
||||
count+=1;consecutive=consecutive+1 if last['idle'] else 0
|
||||
record(dict(**last,poll=count,consecutive_idle=consecutive,elapsed_seconds=clock()-begin))
|
||||
if clock()-begin>limit_seconds:break
|
||||
if consecutive>=3:return dict(status='IDLE_CONFIRMED',uuid=uuid,polls=count,consecutive_idle=consecutive,wall_seconds=clock()-begin,last=last)
|
||||
remaining=limit_seconds-(clock()-begin)
|
||||
if remaining<=0:break
|
||||
sleep(min(10,remaining))
|
||||
return dict(status='IDLE_TIMEOUT',uuid=uuid,polls=count,consecutive_idle=consecutive,wall_seconds=clock()-begin,last=last)
|
||||
|
||||
|
||||
def recorded_wait(root,name,uuid,limit_seconds):
|
||||
path=Path(root)/'idle-gates';path.mkdir(exist_ok=True)
|
||||
def record(row):
|
||||
with (path/(name+'.jsonl')).open('a') as f:f.write(json.dumps(row)+'\n')
|
||||
result=wait_idle(uuid,limit_seconds,record)
|
||||
with (path/(name+'.json')).open('x') as f:json.dump(result,f,indent=2);f.write('\n')
|
||||
return result
|
||||
@@ -0,0 +1,122 @@
|
||||
"""Sequential four-run pilot guard: process-tree RSS/wall plus device VRAM."""
|
||||
import argparse
|
||||
import fcntl
|
||||
import json
|
||||
import os
|
||||
from pathlib import Path
|
||||
import subprocess
|
||||
import sys
|
||||
import threading
|
||||
import time
|
||||
import psutil
|
||||
from freeze_g1_components import sha
|
||||
from s2_recovery_gpu_supervisor import supervise,save,stop_group,disk_usage,tree_usage
|
||||
from s2_pilot_train import select_capacity
|
||||
from s2_gpu_idle_gate import recorded_wait
|
||||
|
||||
|
||||
def parse_device_usage(text):
|
||||
lines=text.strip().splitlines()
|
||||
if len(lines)!=1:raise ValueError('Expected exactly one GPU')
|
||||
fields=lines[0].split(',')
|
||||
if len(fields)!=2:raise ValueError('Malformed GPU usage row')
|
||||
memory,util=map(lambda s:int(s.strip()),fields)
|
||||
if memory<0 or not 0<=util<=100:raise ValueError('Invalid GPU counters')
|
||||
return memory*2**20,util
|
||||
|
||||
|
||||
def device_usage(uuid):
|
||||
r=subprocess.run(['nvidia-smi','--id='+uuid,'--query-gpu=memory.used,utilization.gpu','--format=csv,noheader,nounits'],check=True,capture_output=True,text=True,timeout=2)
|
||||
return parse_device_usage(r.stdout)
|
||||
|
||||
|
||||
def active_training_pids(parent,root,index):
|
||||
ids=[]
|
||||
try:children=psutil.Process(parent).children(recursive=True)
|
||||
except psutil.NoSuchProcess:return ids
|
||||
for p in children:
|
||||
try:
|
||||
cmd=p.cmdline()
|
||||
if str(root) in cmd and (('tools/run_s2_pilot_training_v2.py' in cmd and str(index)==cmd[-1]) or (index==-1 and 'tools/run_s2_pilot_probe_v2.py' in cmd)):ids.append(p.pid)
|
||||
except psutil.NoSuchProcess:pass
|
||||
return ids
|
||||
|
||||
|
||||
def guarded_worker(root,spec,i,uuid,wall_limit,command):
|
||||
run_start=time.monotonic();initial,util=device_usage(uuid)
|
||||
if initial>128*2**20 or util>0:raise RuntimeError('SELECTED_GPU_BUSY_BEFORE_RUN')
|
||||
done=threading.Event();monitor=dict(peak_device_memory_bytes=initial,peak_whole_run_tree_rss_bytes=0,max_poll_gap_seconds=0.,reason=None,samples=0)
|
||||
def watch():
|
||||
previous=time.monotonic()
|
||||
while not done.is_set():
|
||||
now=time.monotonic();monitor['max_poll_gap_seconds']=max(monitor['max_poll_gap_seconds'],now-previous);previous=now
|
||||
try:
|
||||
used,_=device_usage(uuid);monitor['samples']+=1;monitor['peak_device_memory_bytes']=max(monitor['peak_device_memory_bytes'],used)
|
||||
if used>15*2**30:monitor['reason']='DEVICE_VRAM_LIMIT'
|
||||
_,rss,_=tree_usage(os.getpid());monitor['peak_whole_run_tree_rss_bytes']=max(monitor['peak_whole_run_tree_rss_bytes'],rss)
|
||||
if rss>16*2**30:monitor['reason']='WHOLE_RUN_HOST_RSS_LIMIT'
|
||||
if disk_usage(root)>=5*2**30:monitor['reason']='PILOT_DISK_LIMIT'
|
||||
except Exception as exc:monitor['reason']='GPU_MONITOR_FAILED';monitor['error']=str(exc)
|
||||
if monitor['reason']:
|
||||
for pid in active_training_pids(os.getpid(),root,i):stop_group(pid)
|
||||
done.wait(.5)
|
||||
thread=threading.Thread(target=watch,daemon=True);thread.start()
|
||||
try:
|
||||
wall=wall_limit-(time.monotonic()-run_start)-2.5
|
||||
if wall<=0:raise RuntimeError('No budget left before spawn')
|
||||
r=supervise(command,root/'guards',spec['name'],dict(total_cpu_seconds=7200,job_cpu_seconds=wall_limit,job_wall_seconds=wall,
|
||||
tree_rss_bytes=16*2**30,disk_bytes=5*2**30,max_processes=2),poll=.1,gpu_uuid=uuid)
|
||||
finally:
|
||||
done.set();thread.join(timeout=3)
|
||||
total_wall=time.monotonic()-run_start
|
||||
result_path=root/spec['name']/'result.json'
|
||||
result=json.loads(result_path.read_text()) if result_path.exists() else None
|
||||
if monitor['reason']:r.update(status='STOPPED_OR_FAILED',stop_reason=monitor['reason'])
|
||||
if total_wall>wall_limit:r.update(status='STOPPED_OR_FAILED',stop_reason='TOTAL_RUN_WALL_LIMIT')
|
||||
if result and max(result['peak_allocated_bytes'],result['peak_reserved_bytes'])>15*2**30:r.update(status='STOPPED_OR_FAILED',stop_reason='RECORDED_VRAM_LIMIT')
|
||||
r.update(gpu_visible=True,gpu_uuid=uuid,device_monitor=monitor,total_run_wall_seconds=total_wall,
|
||||
result_sha256=sha(result_path) if result else None)
|
||||
save(root/(spec['name']+'-guard-receipt.json'),r)
|
||||
return r,result
|
||||
|
||||
|
||||
def run(root,prior_wall_seconds=0.):
|
||||
with (root/'pilot.lock').open('a') as lock:
|
||||
fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
|
||||
return run_locked(root,prior_wall_seconds)
|
||||
|
||||
|
||||
def run_locked(root,prior_wall_seconds=0.):
|
||||
c=json.loads((root/'training-contract.json').read_text());uuid=c['gpu_uuid']
|
||||
prior_wall_seconds=max(prior_wall_seconds,c.get('probe_guard_wall_seconds',0.))
|
||||
assert [(r['arm'],r['width'],r['seed']) for r in c['runs']]==[('N',128,0),('E',128,0),('N',256,0),('E',256,0)]
|
||||
for p,h in c['code_sha256'].items():assert sha(p)==h,p
|
||||
assert not (root/'pilot-started.json').exists(),'No automatic pilot retry'
|
||||
start=time.monotonic();save(root/'pilot-started.json',dict(started_unix=time.time(),contract_sha256=sha(root/'training-contract.json')))
|
||||
rows=[];scores={};reason=None
|
||||
for i,spec in enumerate(c['runs']):
|
||||
remaining=7200-prior_wall_seconds-(time.monotonic()-start)
|
||||
if remaining<=3:reason='TOTAL_PILOT_WALL_LIMIT';break
|
||||
# Initial readiness was proved by the continuation entry; subsequent
|
||||
# contexts also need a settling window. All waits consume the same budget.
|
||||
if i>0:
|
||||
gate=recorded_wait(root,spec['name'],uuid,min(600,remaining-3))
|
||||
if gate['status']!='IDLE_CONFIRMED':reason='GPU_IDLE_TIMEOUT_BEFORE_RUN';break
|
||||
remaining=7200-prior_wall_seconds-(time.monotonic()-start)
|
||||
if remaining<=3:reason='TOTAL_PILOT_WALL_LIMIT';break
|
||||
command=[sys.executable,'tools/run_s2_pilot_training_v2.py',str(root),str(i)]
|
||||
try:r,result=guarded_worker(root,spec,i,uuid,min(1800,remaining),command)
|
||||
except Exception as exc:reason=str(exc);break
|
||||
rows.append(dict(run=spec,**r))
|
||||
if r['status']!='PASS' or not result or result['status']!='COMPLETE':reason='PILOT_RUN_FAILED_NO_RETRY';break
|
||||
scores[(spec['arm'],spec['width'])]=result['best_validation_score']
|
||||
if prior_wall_seconds+time.monotonic()-start>7200:reason='TOTAL_PILOT_WALL_LIMIT_AFTER_EXIT'
|
||||
selected=select_capacity(scores) if len(scores)==4 and reason is None else None
|
||||
summary=dict(status='FOUR_RUNS_COMPLETE' if selected else 'PILOT_INCOMPLETE',reason=reason,runs=rows,selected_global_width=selected,
|
||||
selection_rule='mean of N/E validation error at each width, ties smaller; all four runs required',
|
||||
pilot_wall_seconds=prior_wall_seconds+time.monotonic()-start,probe_and_setup_wall_seconds=prior_wall_seconds,test_labels_read=False,test_predictions=False,retries=0)
|
||||
save(root/'pilot-complete.json',summary);print(json.dumps(summary),flush=True)
|
||||
|
||||
|
||||
if __name__=='__main__':
|
||||
p=argparse.ArgumentParser();p.add_argument('root',type=Path);a=p.parse_args();run(a.root)
|
||||
@@ -0,0 +1,26 @@
|
||||
import unittest
|
||||
from s2_gpu_idle_gate import wait_idle
|
||||
|
||||
|
||||
class IdleGateTests(unittest.TestCase):
|
||||
def exercise(self,states,limit):
|
||||
now=[0.];events=[];iterator=iter(states);last=[states[-1]]
|
||||
def sample(uuid):
|
||||
value=next(iterator,last[0]);last[0]=value
|
||||
return dict(uuid=uuid,idle=value,occupants=[] if value else [dict(pid=123,user='foreign',memory_MiB='1000')])
|
||||
result=wait_idle('synthetic',limit,events.append,sample=sample,clock=lambda:now[0],sleep=lambda x:now.__setitem__(0,now[0]+x))
|
||||
return result,events
|
||||
def test_busy_then_three_consecutive_free(self):
|
||||
result,events=self.exercise([False,True,False,True,True,True],600)
|
||||
self.assertEqual(result['status'],'IDLE_CONFIRMED');self.assertEqual(result['wall_seconds'],50)
|
||||
self.assertEqual([x['consecutive_idle'] for x in events],[0,1,0,1,2,3])
|
||||
def test_foreign_occupant_times_out_without_killing(self):
|
||||
result,events=self.exercise([False],600)
|
||||
self.assertEqual(result['status'],'IDLE_TIMEOUT');self.assertEqual(result['wall_seconds'],600)
|
||||
self.assertEqual(result['last']['occupants'][0]['pid'],123);self.assertEqual(len(events),61)
|
||||
def test_short_remaining_budget_cannot_start_after_two_idle_polls(self):
|
||||
result,_=self.exercise([True],10)
|
||||
self.assertEqual(result['status'],'IDLE_TIMEOUT')
|
||||
|
||||
|
||||
if __name__=='__main__':unittest.main(verbosity=2)
|
||||
Reference in new issue
Block a user