File size: 3,996 Bytes
d370082 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 | from dataclasses import dataclass
import fcntl
import hashlib
import json
from pathlib import Path
import shutil
import time
from omegaconf import OmegaConf
from rich import print
from refined_v4 import study as runner
from . import REVISION
from .data import ROOT, TASKS
RESULTS = runner.PROJECT / 'results/refined-math-v1'
@dataclass
class Config:
action: str = 'watch'
root: str = str(ROOT)
results: str = str(RESULTS)
capacity: int = 32
interval: int = 30
refresh: int = 120
def prepare(root: Path, results: Path, budget: int):
root.mkdir(parents=True, exist_ok=True)
runner.prepare_results(root, results)
runtime = root / 'runtime-cu126'
if not runtime.exists():
runtime.symlink_to(runner.ROOT / 'runtime-cu126', target_is_directory=True)
source_plot = Path(__file__).parent / '.plotting/plot.py'
(results / '.plotting').mkdir(exist_ok=True)
shutil.copy2(source_plot, results / '.plotting/plot.py')
if (root / 'plan.json').exists():
return runner.read(root / 'plan.json')
cases = []
for model in runner.MODELS:
for task in TASKS:
cases.append(dict(index=len(cases), key=f'{model}__{task}', model=model,
task=task, label=runner.LABELS[task], output=str(root / 'runs' / model / task)))
value = dict(revision=REVISION, root=str(root), results=str(results), tasks=list(TASKS),
models=runner.model_inventory(), cases=cases, sft_tokens=budget, rl_rollouts=65536,
group_size=32, eval_queries=64, eval_samples=8, created=time.time(),
hint_distribution='Uniform1..16 first, then an eligible question and explicit problem-complexity mixture.',
initializer='Existing hinted-v2 initialization; preserved for the first data-only comparison.',
wandb_project='tiny-rl-refined-math-v1', wandb_entity='david-heineman',
format='Native v2 math formats with multiple examples per document; math-v1 episode eligibility.',
partitions=runner.PARTITIONS)
value['signature'] = hashlib.sha256(json.dumps(value, sort_keys=True).encode()).hexdigest()
runner.write(root / 'plan.json', value)
return value
def settings(root: Path, results: Path, budget: int, capacity: int):
return runner.Config(root=str(root), results=str(results), sft_tokens=budget,
capacity=capacity, job_prefix='math1-')
def watch(config: Config):
from .report import update
root, results = Path(config.root), Path(config.results)
root.mkdir(parents=True, exist_ok=True)
pilot = root / 'pilot-2m'
plans = [prepare(pilot, results / 'pilot-2m', 2097152), prepare(root, results, 16777216)]
stages = [settings(pilot, results / 'pilot-2m', 2097152, config.capacity),
settings(root, results, 16777216, config.capacity)]
if config.action == 'plan':
print({key: value for key, value in plans[0].items() if key != 'models'})
return
with (root / 'math-watch.lock').open('a') as lock:
fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
refresh = 0.
while True:
runner.dispatch(stages[0])
pilot_started = all((Path(case['output']) / 'request.json').exists() for case in plans[0]['cases'])
if pilot_started:
runner.dispatch(stages[1])
complete = [sum((Path(case['output']) / 'complete.json').exists() for case in plan['cases']) for plan in plans]
print(f'[cyan]Math-v1: pilot {complete[0]}/16 complete; full {complete[1]}/16 complete[/cyan]', flush=True)
if time.monotonic() - refresh >= config.refresh:
for stage in stages:
update(Path(stage.root), Path(stage.results))
refresh = time.monotonic()
if complete == [16, 16]:
return
time.sleep(config.interval)
if __name__ == '__main__':
watch(OmegaConf.to_object(OmegaConf.merge(OmegaConf.structured(Config), OmegaConf.from_cli())))
|