source/project/scripts/resume_step3000_api_eval.py

#!/usr/bin/env python3
"""Resume only API mechanics judges; immutable videos and original scoring protocol."""
from __future__ import annotations
import argparse, collections, concurrent.futures, datetime, fcntl, http.client, importlib.util
import json, os, random, sys, threading, time, urllib.error, urllib.request, zipfile
from pathlib import Path

ROOT=Path(__file__).resolve().parents[1]
spec=importlib.util.spec_from_file_location('step3000_batch',ROOT/'scripts/benchmark_new_step3000.py')
batch=importlib.util.module_from_spec(spec);spec.loader.exec_module(batch)
BENCH=batch.BENCH;SOURCE=batch.SOURCE
spec=importlib.util.spec_from_file_location('step3000_original_evaluator',SOURCE/'stateplay_eval.py')
base=importlib.util.module_from_spec(spec);sys.modules[spec.name]=base;spec.loader.exec_module(base)
MODELS={'gemini':('gemini','gemini-3.1-pro-preview'),'gpt61':('gpt','gpt-6.1-sol'),'gpt55':('gpt','gpt-5.5')}
SLOTS=threading.BoundedSemaphore(32)
STATUS_LOCK=threading.Lock();LANE_STATUS={}

class AuthenticationFailure(RuntimeError):pass

def report(lane,stage,**info):
 with STATUS_LOCK:
  LANE_STATUS[lane]=dict(stage=stage,time=batch.now(),**info)
  batch.write_json(BENCH/'api_resume_status.json',dict(time=batch.now(),total_max_inflight_requests=32,models={k:v[1] for k,v in MODELS.items()},lanes=LANE_STATUS))
 print(batch.now(),lane,stage,json.dumps(info),flush=True)

original_parse=base.parse_json_object

def parse_json_object(text):
 """Only tolerate literal control characters inside JSON strings, not invalid labels."""
 try:return original_parse(text)
 except json.JSONDecodeError:
  stripped=text.strip()
  if stripped.startswith('```'):
   stripped=stripped.split('\n',1)[-1].rsplit('```',1)[0].strip()
  try:value=json.loads(stripped,strict=False)
  except json.JSONDecodeError:
   start,end=stripped.find('{'),stripped.rfind('}')
   if start<0 or end<=start:raise ValueError('Judge did not return a complete JSON object')
   value=json.loads(stripped[start:end+1],strict=False)
  if isinstance(value,list) and len(value)==1 and isinstance(value[0],dict):value=value[0]
  if not isinstance(value,dict):raise ValueError('Expected one JSON object')
  value['predicted_state']=base.normalized_label(str(value.get('predicted_state','')))
  return value


def http_json(url,payload,headers,timeout,retries):
 data=json.dumps(payload).encode('utf-8')
 for attempt in range(retries+1):
  try:
   req=urllib.request.Request(url,data=data,headers={'Content-Type':'application/json',**headers},method='POST')
   # All three independent judge lanes share a total of 32 HTTP request slots.
   with SLOTS:
    with urllib.request.urlopen(req,timeout=timeout) as response:
     value=json.loads(response.read().decode('utf-8'),strict=False)
   if not isinstance(value,dict):raise ValueError('Expected an API response object')
   return value
  except urllib.error.HTTPError as error:
   code=error.code;error.close()
   if code in {401,403}:raise AuthenticationFailure(f'API authentication/permission failed with HTTP {code}')
   if attempt==retries:raise RuntimeError(f'API HTTP {code} after {retries+1} attempts')
   delay=min(10*(2**attempt),90) if code in {429,500,502,503,504,524} else min(2**attempt,8)
   print(f'API HTTP {code}; retry {attempt+1}/{retries} in {delay}s',flush=True)
  except (urllib.error.URLError,http.client.HTTPException,ConnectionError,TimeoutError,json.JSONDecodeError) as error:
   if attempt==retries:raise RuntimeError(f'API transport failed after {retries+1} attempts: {type(error).__name__}')
   delay=min(5*(2**attempt),60);print(f'API transport {type(error).__name__}; retry {attempt+1}/{retries} in {delay}s',flush=True)
  time.sleep(delay)
 raise AssertionError('unreachable')

base.parse_json_object=parse_json_object;base.http_json=http_json


def make_args(ds,lane,credentials):
 judge,model=MODELS[lane]
 args=base.build_parser().parse_args(['--manifest',str(BENCH/ds/'manifest_mechanics.csv'),'--visual-metrics','--mechanics',judge,'--mechanics-runs','3','--mechanics-workers','32','--gemini-transport','openai','--gemini-model','gemini-3.1-pro-preview','--gpt-model',model if judge=='gpt' else 'gpt-6.1-sol','--gemini-frames','24','--gpt-frames','24','--judge-image-max-side','512','--judge-max-completion-tokens','4096','--api-timeout','300','--api-retries','5','--judge-retries','3'])
 args.gpt_api_key=credentials['STATEPLAY_GPT_API_KEY'];args.gemini_api_key=credentials['STATEPLAY_GEMINI_API_KEY'];args.gpt_base_url=args.gemini_base_url='https://newapi.dragon3api.com/v1'
 return args


def identity(sample,judge,model,run,args):
 stat=sample.prediction.stat()
 return dict(judge=judge,model=model,transport=args.gemini_transport if judge=='gemini' else 'openai',run=run,id=sample.sample_id,ground_truth=sample.label,prediction=str(sample.prediction),prediction_size=stat.st_size,prediction_mtime_ns=stat.st_mtime_ns,frame_count=args.gemini_frames if judge=='gemini' else args.gpt_frames,image_max_side=args.judge_image_max_side,protocol_version=base.MECHANICS_PROTOCOL_VERSION)


def evaluate_one(sample,item,args,judge):
 method=base.judge_gemini if judge=='gemini' else base.judge_gpt
 for attempt in range(args.judge_retries+1):
  try:
   answer=method(sample,args)
   return dict(**item,predicted_state=answer['predicted_state'],correct=answer['predicted_state']==sample.label,confidence=answer.get('confidence'),evidence=answer.get('evidence'),reason=answer.get('reason'))
  except AuthenticationFailure:raise
  except (RuntimeError,ValueError) as error:
   if attempt==args.judge_retries:raise
   print(f"[{item['model']} run {item['run']}] retry {attempt+1} {sample.sample_id}: {type(error).__name__}",flush=True)
   time.sleep(min(2**attempt,8))


def evaluate_dataset(ds,lane,credentials):
 args=make_args(ds,lane,credentials);judge,model=MODELS[lane];folder=BENCH/ds;output=folder/f'results_{lane}.json';cache_path=folder/f'results_{lane}.mechanics.jsonl'
 samples=base.load_manifest(args.manifest.resolve(),require_reference=False)
 # Preserve all 20 successful original Gemini requests using the exact original cache identity.
 legacy=folder/'results_gemini_gpt61.mechanics.jsonl'
 if not cache_path.exists() and legacy.exists():
  selected=[row for row in base.load_cache(legacy).values() if row['judge']==judge and row['model']==model]
  for row in selected:base.append_cache(cache_path,row)
 if output.exists():
  d=json.loads(output.read_text());v=d['mechanics_fidelity'][judge]
  if len(v['records'])==len(samples)*3 and all(r['model']==model for r in v['records']):
   report(lane,'dataset_already_complete',dataset=ds,mean_percent=v['mean_percent']);return d
 cache=base.load_cache(cache_path);records=[];accuracies=[]
 for run in range(1,4):
  wave=0
  while True:
   results=[None]*len(samples);pending=[]
   for position,sample in enumerate(samples):
    item=identity(sample,judge,model,run,args);key=base.cache_key(item)
    if key in cache:results[position]=cache[key]
    else:pending.append((position,sample,item,key))
   if not pending:break
   wave+=1;report(lane,'judging',dataset=ds,run=run,wave=wave,completed=len(samples)-len(pending),expected=len(samples),pending=len(pending))
   failures=[];auth_failure=None
   with concurrent.futures.ThreadPoolExecutor(max_workers=min(32,len(pending)),thread_name_prefix=lane) as pool:
    futures={pool.submit(evaluate_one,sample,item,args,judge):(position,item,key) for position,sample,item,key in pending}
    for future in concurrent.futures.as_completed(futures):
     position,item,key=futures[future]
     try:row=future.result()
     except AuthenticationFailure as error:
      auth_failure=error;failures.append(item['id']);continue
     except (RuntimeError,ValueError) as error:
      failures.append(item['id']);print(f"[{model}] unresolved {item['id']}: {type(error).__name__}",flush=True);continue
     # Save successes even when another request has failed; no results are discarded.
     results[position]=row;cache[key]=row;base.append_cache(cache_path,row)
     done=sum(row is not None for row in results);report(lane,'judging',dataset=ds,run=run,wave=wave,completed=done,expected=len(samples))
   if auth_failure:raise auth_failure
   if failures:
    report(lane,'retrying_missing',dataset=ds,run=run,missing=len(failures),pause_seconds=60);time.sleep(60)
  completed=[cache[base.cache_key(identity(sample,judge,model,run,args))] for sample in samples]
  check_count=len(completed);batch.check(check_count==len(samples),'Incomplete round must not be scored')
  accuracies.append(100*sum(r['correct'] for r in completed)/len(samples));records.extend(completed)
  report(lane,'round_complete',dataset=ds,run=run,accuracy_percent=accuracies[-1])
 import statistics
 result=dict(manifest=str(args.manifest.resolve()),protocol=dict(visual_alignment='min frame count; reference center-crop + bicubic resize to prediction/target size',ssim=None,lpips=None,mechanics_runs=3,mechanics_workers=32,total_http_concurrency_across_judges=32,gemini_transport='openai' if judge=='gemini' else None,gpt_uniform_frames=24 if judge=='gpt' else None,judge_image_max_side=512,judge_max_completion_tokens=4096,mechanics_protocol_version=base.MECHANICS_PROTOCOL_VERSION,json_control_characters_tolerated=True,all_expected_requests_required_before_scoring=True),mechanics_fidelity={judge:dict(judge=judge,model=model,n_runs=3,workers=32,run_accuracies_percent=accuracies,mean_percent=statistics.fmean(accuracies),std_percent=statistics.pstdev(accuracies),records=records)})
 batch.write_json(output,result);report(lane,'dataset_complete',dataset=ds,mean_percent=result['mechanics_fidelity'][judge]['mean_percent']);return result


def lane_work(lane,credentials):
 try:
  for ds in ['test200','test100']:evaluate_dataset(ds,lane,credentials)
  report(lane,'complete')
 except BaseException as error:
  report(lane,'failed',error_type=type(error).__name__);raise


def finish():
 combined={};table=[]
 for ds in ['test100','test200']:
  p=BENCH/ds;combined[ds]={key:json.loads(path.read_text()) for key,path in [('visual',p/'results_visual.json'),('hud',p/'hud_results/hud_metrics_summary.json')]+[(lane,p/f'results_{lane}.json') for lane in MODELS]}
  metrics=combined[ds];visual=metrics['visual']['visual']['aggregate'];row=dict(dataset=ds,ssim=visual['ssim']['sample_mean'],lpips=visual['lpips']['sample_mean'],rvsa=metrics['hud']['aggregate']['rvsa_low'])
  for lane,(judge,model) in MODELS.items():
   v=metrics[lane]['mechanics_fidelity'][judge];row[lane+'_mean_percent']=v['mean_percent'];row[lane+'_std_percent']=v['std_percent']
  table.append(row)
 batch.write_json(BENCH/'all_results.json',dict(model=json.loads((BENCH/'model_identity.json').read_text()),datasets=combined));batch.write_csv(BENCH/'final_metrics_summary.csv',table)
 lines=['# 新step-3000完整评测结果','', '| 测试集 | SSIM↑ | LPIPS↓ | RVSA↑ | Gemini (%) | GPT-6.1-sol (%) | GPT-5.5 (%) |','|---|---:|---:|---:|---:|---:|---:|']
 for row in table:lines.append(f"| {row['dataset']} | {row['ssim']:.6f} | {row['lpips']:.6f} | {row['rvsa']:.6f} | {row['gemini_mean_percent']:.3f} ± {row['gemini_std_percent']:.3f} | {row['gpt61_mean_percent']:.3f} ± {row['gpt61_std_percent']:.3f} | {row['gpt55_mean_percent']:.3f} ± {row['gpt55_std_percent']:.3f} |")
 lines+=['','权重仅为/private/step-3000.safetensors，SHA256 3e14d04dd0cd4b7cf5d6dfb10c70a554411800240581db38dbd94aa494e5398e。三个裁判各做3轮，完整样本才计分；API总并发上限32，提示、24张抽帧、512最大边和4096最大输出tokens与原评测相同。仅容忍JSON字符串中的未转义控制字符，标签仍严格限制为原五类。成功缓存保留，失败请求不计作错误样本。100条是200条精确子集。']
 (BENCH/'FINAL_REPORT_ZH.md').write_text('\n'.join(lines)+'\n')
 archive=BENCH.parent/(batch.RUN_NAME+'_results.zip')
 with zipfile.ZipFile(archive,'w',zipfile.ZIP_DEFLATED) as z:
  for p in BENCH.rglob('*'):
   if p.is_file() and not any(x in p.parts for x in ['generated','source_snapshot','quarantine']) and not p.name.startswith('launcher') and p.suffix!='.tmp':z.write(p,str(p.relative_to(BENCH.parent)))
 batch.status('complete',results_archive=str(archive),api_models=[v[1] for v in MODELS.values()],total_http_concurrency=32)


def run(args,lock_already_held=False,credentials=None):
 lock=None
 if not lock_already_held:
  lock=(BENCH/'.run.lock').open('a')
  try:fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
  except BlockingIOError:raise RuntimeError('A batch controller is already running')
 pipe=None
 try:
  model=json.loads((BENCH/'model_identity.json').read_text());batch.check(model['checkpoint_sha256']==batch.MODEL_SHA and model['checkpoint']==str(batch.CHECKPOINT),'Wrong model identity')
  batch.check(batch.sha(batch.CHECKPOINT)==batch.MODEL_SHA,'Current model bytes differ')
  for ds,count in [('test200',200),('test100',100)]:
   verification=json.loads((BENCH/ds/f'verification_{batch.RUN_NAME}.json').read_text());batch.check(verification['verified_videos']==count,'Wrong dataset size')
   batch.check((BENCH/ds/'results_visual.json').exists() and (BENCH/ds/'hud_results/hud_metrics_summary.json').exists(),'Local results are missing')
  values,ready,pipe=batch.credential_channel(credentials if credentials is not None else batch.load_credentials(args.credential_file));batch.status('api_resume_waiting_for_credentials',models=[v[1] for v in MODELS.values()],concurrency=32)
  ready.wait();batch.status('api_judges_running',models=[v[1] for v in MODELS.values()],total_http_concurrency=32)
  errors=[]
  with concurrent.futures.ThreadPoolExecutor(max_workers=3,thread_name_prefix='judge-lane') as pool:
   futures={pool.submit(lane_work,lane,values):lane for lane in MODELS}
   for future in concurrent.futures.as_completed(futures):
    try:future.result()
    except BaseException as error:errors.append((futures[future],type(error).__name__))
  if errors:raise RuntimeError('Independent judge lanes failed: '+str(errors))
  finish()
 except BaseException as error:
  batch.status('failed',error_type=type(error).__name__,details=str(error));raise
 finally:
  if pipe is not None and pipe.exists():pipe.unlink()
  if lock is not None:lock.close()


def main():
 parser=argparse.ArgumentParser(description=__doc__);parser.add_argument('--detach',action='store_true');parser.add_argument('--credential-file',type=Path);args=parser.parse_args();os.umask(0o077)
 if args.detach:
  log=BENCH/'logs/api_resume_launcher.log';cmd=[sys.executable,'-u',str(Path(__file__).resolve())]
  if args.credential_file:cmd+=['--credential-file',str(args.credential_file.resolve())]
  with log.open('ab') as f:p=__import__('subprocess').Popen(cmd,cwd=ROOT,stdin=__import__('subprocess').DEVNULL,stdout=f,stderr=__import__('subprocess').STDOUT,start_new_session=True)
  (BENCH/'launcher.pid').write_text(str(p.pid)+'\n');time.sleep(2);batch.check(p.poll() is None,'API resume launcher exited; see '+str(log));print(json.dumps(dict(pid=p.pid,log=str(log),models=[v[1] for v in MODELS.values()],max_inflight_http_requests=32),indent=2));return
 run(args)

if __name__=='__main__':main()