| """Exact PP1 law using q emitted by the real GPU producer (synthetic only).""" |
| from __future__ import annotations |
|
|
| import argparse |
| import hashlib |
| import importlib |
| import json |
| import os |
| from fractions import Fraction |
| from pathlib import Path |
|
|
| import torch |
|
|
| from offload.dspark.probabilistic.reference import Categorical, acceptance, residual |
| from offload.dspark.worker.sampled import MASS, Top32 |
| from offload.dspark.worker.tools.pp3_probe import UUID |
|
|
|
|
| def main() -> None: |
| p = argparse.ArgumentParser(description=__doc__) |
| p.add_argument('--out',type=Path,required=True) |
| a=p.parse_args() |
| if os.environ.get('CUDA_VISIBLE_DEVICES') != UUID: |
| raise RuntimeError('PP3 solo UUID required') |
| import fcntl |
| fcntl.flock(int(os.environ['PP3_LOCK_FD']),fcntl.LOCK_EX|fcntl.LOCK_NB) |
| if str(torch.cuda.get_device_properties(0).uuid).removeprefix('GPU-') != UUID.removeprefix('GPU-'): |
| raise RuntimeError('wrong GPU') |
| torch.set_num_threads(8) |
| torch.set_num_interop_threads(1) |
| dev=torch.device('cuda:0') |
| sampler=Top32(1,64,dev) |
| rows=[] |
| for name, support, weights in ( |
| ('full',[40,41,42,43],[4.,1.,1.,2.]), |
| ('sparse',[40,41],[1.,1.]), |
| ('disjoint',[0,1,2,3],[4.,1.,1.,2.]), |
| ('equal',[40,41,42,43],[1.,2.,2.,3.])): |
| x=torch.full((1,64),-torch.inf,device=dev) |
| x[0,support]=torch.tensor(weights,device=dev).log()*.7 |
| token,ids,q=sampler(x,torch.tensor([[4103,1,32768]],device=dev),torch.tensor([.7],device=dev)) |
| if sampler.error.any().item(): |
| raise AssertionError('GPU producer refused bridge input') |
| counts=(q[0].double()*MASS).long().cpu().tolist() |
| support_ids=ids[0].cpu().tolist() |
| assert sum(counts)==MASS |
| actual_q=Categorical(64,support_ids,counts) |
| target=Categorical(64,[40,41,42,43],[1,2,2,3]) |
| accepted=[actual_q.prob(i)*acceptance(target,actual_q,i) if actual_q.prob(i) else Fraction() |
| for i in range(64)] |
| reject=1-sum(accepted) |
| correction=residual(target,actual_q) if reject else None |
| law=[accepted[i]+(reject*correction.prob(i) if correction is not None else 0) for i in range(64)] |
| assert law==[target.prob(i) for i in range(64)] |
| assert actual_q.prob(int(token.item()))>0 |
| rows.append({'case':name,'support_ids':support_ids,'counts':counts,'exact_law':True}) |
| receipt={'schema':'pp3.gpu-q-pp1-oracle.v1','classification':'MEASURED actual GPU q, exact rational CPU law', |
| 'source_sha256':hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),'cases':rows} |
| for module_name in ('offload.dspark.worker.sampled', 'offload.dspark.probabilistic.reference'): |
| module_path=importlib.import_module(module_name).__file__ |
| if module_path is None: |
| raise RuntimeError('oracle/producer source path missing') |
| receipt[module_name+'_sha256']=hashlib.sha256(Path(module_path).read_bytes()).hexdigest() |
| a.out.write_text(json.dumps(receipt,indent=2)+'\n') |
| print('PASS: actual GPU q recovers full target exactly in PP1 oracle (four synthetic cases)',flush=True) |
|
|
|
|
| if __name__ == '__main__': |
| main() |
|
|
