code-Movens-Zero_train / scale /script /train_reference_baseline_fresh.py
Tinghong-Ye22's picture
Add files using upload-large-folder tool
13cc860 verified
Raw History Blame Contribute Delete
11.2 kB
"""Official Omni + random normal-size expert; exact brother training worker.
Storage/cache paths and initialization provenance are isolated. No training
implementation is patched. Effective LRs, batch/accumulation and loss follow
the brother's saved 89k configuration. This is NOT replay of his warmstart.
"""
import argparse
import copy
import datetime
import fcntl
import hashlib
import json
import os
from pathlib import Path
import shutil
import signal
import subprocess
import sys
import threading
import time
ROOT = Path(__file__).resolve().parents[2]
WS = ROOT.parents[2]
LOCAL = Path('/mnt/workspace/Movens-Zero_train/scale/baseline_reference_20261008')
SOURCE = LOCAL/'source'
BRO = WS/'buwendong/user/code/OmniHM-Zero_train_40d'
REF = WS/'buwendong/user/tmp/train_dashboard_20260930/case_gallery/ckpt_89000__root40_v5/snapshot'
ARCHIVE = ROOT/'scale/baseline_reference_20261008'
PYTHON = ROOT.parent/'Psi0-yth-v26/.venv-psi/bin/python'
GPU = 3
PORT = 29639
sys.path.insert(0,str(ROOT/'scale/script'))
from dlc_local_storage import copy_verified, save
import preview_gpu_lease as gpu
def digest(path):
return hashlib.sha256(path.read_bytes()).hexdigest()
def prepare():
config_path=LOCAL/'fresh_config.json'
if config_path.exists():
return config_path
if not (SOURCE/'train_vlm.py').exists():
raise RuntimeError('First run scale/audits/training_baseline_20261008/audit.py')
source_hashes=json.loads((ROOT/'scale/audits/training_baseline_20261008/source_hashes.json').read_text())
for rel,expected in source_hashes.items():
if digest(SOURCE/rel)!=expected:
raise RuntimeError(f'Frozen reference code was changed: {rel}')
original=json.loads((REF/'run_config.json').read_text())
cfg=copy.deepcopy(original)
cfg.update(exp='baseline_reference_official_omni_random_action',
timestamp=datetime.datetime.now().strftime('%y%m%d%H%M%S'),auto_tag_run=False,
initial_action_checkpoint=None,warmstart_checkpoint=None,warmstart_provenance={},
resume_provenance={},resume_transition_provenance={},world_size_expected=1)
cfg['absolute_height_stage']=dict(start_step=0,end_step=100000,lr_scale=0.1,
scheduler_policy='constant_at_reference_effective_lr',
source_checkpoint='NONE:official_omni_random_action_fresh_initialization')
cfg['train'].update(resume_from_checkpoint=None,resume_load_full_state=False,skip_resumed_steps=False,
output_dir=str(LOCAL/'runs'),max_checkpoints_to_keep=2)
cfg['model']['pretrained_action_header_path']=None
cfg['log']['report_to']='none'
cfg['wandb'].update(id=None,name=None,resume='never',project='Movens-reference-baseline-fresh')
# Keep all data sources/metadata and their logical fingerprint identical.
# Only copy and relocate writable caches; do not write to shared originals.
old_cache=BRO/'training_set/cache/records'
new_cache=LOCAL/'cache/records'
shutil.copytree(old_cache,new_cache,dirs_exist_ok=True)
relocations={}
for field in ['text','audio','audio_command','video_clip']:
data=cfg['data'][field]
sources=json.loads(data['sources_json'])
for item in [data]+sources:
old=item.get('records_cache_dir')
if old:
new=str(new_cache/Path(old).relative_to(old_cache))
item['records_cache_dir']=new;relocations[old]=new
data['sources_json']=json.dumps(sources)
# Never retain the brother's machine-local cache/preparation output paths.
cfg['train']['jerk_kinematics_path']=str(SOURCE/'assets/robots/g1/g1_kinematics.json')
if cfg['train'].get('absolute_height_startup_seed_path'):
seed=LOCAL/'startup_pose_motion36.json'
origin=Path(original['train']['absolute_height_startup_seed_path'])
if origin.exists():
shutil.copy2(origin,seed)
else:
# Auxiliary loss is disabled, so this path is deliberately unused.
seed=LOCAL/'UNUSED_startup_pose_motion36.json'
cfg['train']['absolute_height_startup_seed_path']=str(seed)
shutil.copy2(REF/'dataset_statistics.json',LOCAL/'dataset_statistics.json')
cfg['data']['transform']['field']['stat_path']=str(LOCAL/'dataset_statistics.json')
assert (cfg['model']['hidden_dim'],cfg['model']['num_blocks'],cfg['model']['nhead'])==(1536,6,24)
t=cfg['train']
assert (t['train_batch_size'],t['gradient_accumulation_steps'],t['max_training_steps'])==(32,12,100000)
assert t['absolute_height_loss_weight']==t['jerk_loss_weight']==0
assert t['cfg_condition_dropout_prob']==0.1 and t['warmup_steps']==0
assert abs(t['learning_rate']*.1-1e-6)<1e-15
assert abs(cfg['model']['lang_backbone_lr']*.1-1e-7)<1e-15
save(config_path,cfg)
save(LOCAL/'reference_config.json',original)
save(LOCAL/'provenance.json',dict(reference_config=str(REF/'run_config.json'),
reference_config_sha256=digest(REF/'run_config.json'),source_hashes=source_hashes,
initialization='official Omni + random normal-size action; optimizer/scheduler/RNG/data cursor fresh',
batch=32,accumulation=12,global_batch=384,effective_action_lr=1e-6,effective_omni_lr=1e-7,
cache_relocations=relocations,storage_changes='independent outputs; latest two checkpoints',
code_changes='NONE: original byte-identical train_vlm.py and src execute --worker-config',
optimizer_history_reproduced=False,reference_training_history_reproduced=False))
for name in ['fresh_config.json','reference_config.json','provenance.json','dataset_statistics.json']:
copy_verified(LOCAL/name,ARCHIVE/name)
return config_path
def backup_once():
for run in (LOCAL/'runs/posttrain_v1_vlm').glob('*'):
if not run.is_dir():continue
destination=ARCHIVE/'runs'/run.name
for name in ['run_config.json','dataset_statistics.json','metrics.jsonl']:
if (run/name).is_file():
if name == 'metrics.jsonl':
# copy_verified deliberately never replaces existing data;
# live metrics need a verified new snapshot, then atomic publish.
destination.mkdir(parents=True,exist_ok=True)
temporary=destination/f'.metrics.snapshot.{os.getpid()}'
if temporary.exists():
temporary.unlink() # only our own unfinished log snapshot
copy_verified(run/name,temporary,append_only=True)
os.replace(temporary,destination/name)
else:
copy_verified(run/name,destination/name)
markers=sorted((run/'checkpoints').glob('ckpt_*/checkpoint_complete.json'),key=lambda p:int(p.parent.name[5:]))
for marker in markers:
ckpt=marker.parent;dest=destination/'checkpoints'/ckpt.name
if (dest/'BACKUP_COMPLETE.json').exists():continue
files={}
for src in sorted(ckpt.iterdir()):
if src.is_file():files[src.name]=copy_verified(src,dest/src.name)
save(dest/'BACKUP_COMPLETE.json',dict(owner='reference-baseline-fresh-v1',step=int(ckpt.name[5:]),files=files))
save(ARCHIVE/'latest.json',dict(run=str(destination),step=int(ckpt.name[5:]),checkpoint=str(dest)))
completed=sorted((destination/'checkpoints').glob('ckpt_*/BACKUP_COMPLETE.json'),key=lambda p:int(p.parent.name[5:]))
for marker in completed[:-2]:
if json.loads(marker.read_text()).get('owner')!='reference-baseline-fresh-v1':
raise RuntimeError('Refusing to prune unowned backup')
target=marker.parent.resolve()
if not target.is_relative_to(ARCHIVE.resolve()) or not target.name.startswith('ckpt_'):
raise RuntimeError('Unsafe backup pruning target')
shutil.rmtree(target)
def main():
parser=argparse.ArgumentParser();parser.add_argument('--prepare-only',action='store_true')
args=parser.parse_args();LOCAL.mkdir(parents=True,exist_ok=True)
with (LOCAL/'supervisor.lock').open('a') as lock:
fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
config=prepare()
if args.prepare_only:
print(config,flush=True);return
if (LOCAL/'launch_state.json').exists():
raise RuntimeError('A launch was already attempted. Inspect it; never silently restart from zero.')
# Physical GPU 3 has a separate existing keeper; do not touch GPU 2's previews.
gpu.PREVIEW_GPU=GPU;gpu.KEEPER_PREFIX=f'/tmp/movens_scale_preview_gpu{GPU}_keeper'
gpu.PID_FILES=(Path(gpu.KEEPER_PREFIX+'_supervisor.pid'),Path(gpu.KEEPER_PREFIX+'.lock'))
gpu.gpu_processes.__defaults__=(GPU,)
def interrupt(sig,frame):raise KeyboardInterrupt(sig)
signal.signal(signal.SIGTERM,interrupt);signal.signal(signal.SIGINT,interrupt)
stop=threading.Event()
def backups():
while not stop.is_set():
try:backup_once()
except Exception as e:save(LOCAL/'backup_status.json',dict(error=repr(e),time=time.time()))
else:save(LOCAL/'backup_status.json',dict(status='ok',time=time.time()))
stop.wait(60)
thread=threading.Thread(target=backups,daemon=True)
with gpu.lease_gpu2():
env=os.environ.copy()
env.update(CUDA_VISIBLE_DEVICES=str(GPU),PSI_DATALOADER_PIN_MEMORY='1',PSI_DATALOADER_PREFETCH_FACTOR='2',
PSI_WORKER_TORCH_THREADS='1',OPENBLAS_NUM_THREADS='1',MKL_NUM_THREADS='1',
PSI_TRAIN_TORCH_THREADS='8',PSI_VIDEO_FAST_SEEK='1',
PSI_VIDEO_LOCAL_CACHE_ROOT=str(LOCAL/'cache/video_clips'),
PSI_VIDEO_CACHE_MIN_FREE_GIB='256',PSI_VIDEO_CACHE_MAX_SOURCE_GIB='2',
HF_HUB_OFFLINE='1',PYTHONDONTWRITEBYTECODE='1',WANDB_MODE='disabled')
command=[str(PYTHON),'-B','-u','-m','torch.distributed.run','--nproc_per_node=1',
f'--master_port={PORT}',str(SOURCE/'train_vlm_shared_gpu.py'),'--worker-config',str(config)]
with (LOCAL/'training.log').open('ab') as log:
child=subprocess.Popen(command,cwd=SOURCE,env=env,stdout=log,stderr=subprocess.STDOUT,start_new_session=True)
save(LOCAL/'launch_state.json',dict(pid=child.pid,supervisor_pid=os.getpid(),time=time.time(),
command=command,gpu=GPU,status='started',initialization='official_omni_random_action'))
thread.start()
try:
while child.poll() is None:time.sleep(5)
save(LOCAL/'exit_status.json',dict(returncode=child.returncode,time=time.time()))
if child.returncode:raise RuntimeError(f'Training failed with {child.returncode}; inspect training.log')
finally:
if child.poll() is None:
os.killpg(child.pid,signal.SIGTERM)
try:child.wait(timeout=45)
except subprocess.TimeoutExpired:os.killpg(child.pid,signal.SIGKILL);child.wait()
stop.set();thread.join(timeout=5)
if __name__=='__main__':main()