Blame for python-balanced-20261007/provenance/generate.py

"""Extend both Python arms without changing source rows or selecting by reward."""
from pathlib import Path
import os,sys,json,hashlib,fcntl,copy,argparse
from concurrent.futures import ThreadPoolExecutor, wait, FIRST_COMPLETED
from dataclasses import replace
BASE=Path(__file__).resolve().parent;ROOT=BASE.parents[2]
os.environ.setdefault('CGD_ARTIFACTS',str(ROOT/'artifacts'))
sys.path.insert(0,str(BASE/'runtime'))
from code_search import data,harness,sft,vllm,colgrep
from code_search.harness.manifest import SamplingSpec
from code_search.sft.export import export_sft_trace
from code_search.colgrep import run as colgrep_run
from code_search.terminal.sandbox import Terminal

MODEL='deepseek-ai/DeepSeek-V4-Flash-0731'
def validate_messages(row):
 pending=set();seen=set()
 for m in row['messages']:
  if m['role']=='assistant':
   if pending:raise ValueError('Missing tool reply')
   calls=m.get('tool_calls') or []
   if len(calls)>5:raise ValueError('Tool call limit')
   for c in calls:
    cid=c['id'];a=json.loads(c['function']['arguments'])
    if not isinstance(a,dict) or not cid or cid in seen:raise ValueError('Malformed tool call')
    pending.add(cid);seen.add(cid)
  elif m['role']=='tool':
   if m.get('tool_call_id') not in pending:raise ValueError('Orphan tool reply')
   pending.remove(m['tool_call_id'])
 if pending:raise ValueError('Missing final tool reply')

def manifests(arm):
 student=json.loads((BASE/f'{arm}-student-protocol.json').read_text())
 common=dict(model_id=MODEL,parser_profile='deepseek_v4',search_tool=arm,max_turns=10,max_retries=6,remind_before_end=2,finish_turns=2,rescue_policy='force_then_harvest',sampling=SamplingSpec(temperature=.3,max_tokens=1024),colgrep=harness.ColgrepSpec(model='lightonai/LateOn-Code-edge',top_k=10,force_cpu=True,no_update=True))
 export=harness.build_structured_manifest(**common)
 assert export.system_prompt==student['system_prompt'] and list(export.tool_schema)==student['tools']
 teacher=harness.build_structured_manifest(**common,colgrep_only=arm=='colgrep',extra={'purpose':'python_unfiltered_count_topup','runtime':str(BASE/'runtime'),'thinking':False})
 return teacher,export

def rolling_results(fn, items, needed, workers):
 # Bound successful + in-flight rows by the remaining quota, avoiding overshoot.
 cursor=0; accepted=0; failures=0
 with ThreadPoolExecutor(max_workers=workers) as pool:
  active=set()
  while accepted<needed:
   while cursor<len(items) and len(active)<workers and accepted+len(active)<needed:
    active.add(pool.submit(fn,items[cursor]));cursor+=1
   if not active:break
   ready,active=wait(active,return_when=FIRST_COMPLETED)
   for future in ready:
    row,error=future.result()
    if error:failures+=1
    else:accepted+=1;failures=0
    yield row,error
    if failures>=8:raise RuntimeError('Eight consecutive generation failures; stopping')

def run_arm(arm,url,pilot=False,workers=48):
 prep=json.loads((BASE/'preparation.json').read_text()); target=prep['baseline'][arm]['additional_needed']
 state=BASE/arm;state.mkdir(exist_ok=True);lock=(state/'generation.lock').open('w');fcntl.flock(lock,fcntl.LOCK_EX|fcntl.LOCK_NB)
 teacher_manifest,student_manifest=manifests(arm)
 manifestfile=state/'generation-manifest.json'
 if manifestfile.exists():assert json.loads(manifestfile.read_text())==teacher_manifest.to_dict()
 else:teacher_manifest.write(manifestfile)
 rawfile=state/'new-raw.jsonl';done={}
 if rawfile.exists():
  for line in rawfile.open():
   r=json.loads(line);assert r['instance_id'] not in done;done[r['instance_id']]=r
 candidates=[data.Instance.from_json(json.loads(l)) for l in (BASE/'tasks.jsonl').open()]
 assert hashlib.sha256((BASE/'tasks.jsonl').read_bytes()).hexdigest()==prep['candidate_tasks_sha256']
 pending=[r for r in candidates if r.instance_id not in done]
 llm=vllm.Client(model=MODEL,base_url=url,temperature=.3,max_tokens=1024,chat_template_kwargs={'enable_thinking':False},max_retries=4)
 vllm.assert_tool_calls(llm,list(teacher_manifest.tool_schema),system_prompt=teacher_manifest.system_prompt)
 colgrep_run._LOCAL_HF=os.environ.get('HF_HOME',colgrep_run._LOCAL_HF)
 def one(inst):
  try:
   outcome=harness.run_session(llm,instance_id=inst.instance_id,task=inst.problem_statement,repo=inst.checkout_dir,manifest=teacher_manifest)
   if outcome.end_reason in {'llm_error','context_overflow'}:raise RuntimeError(outcome.error or outcome.end_reason)
   row=export_sft_trace(outcome,manifest=teacher_manifest,gold={'files':sorted(inst.gold_files)})
   score=sft.score_instance(inst.instance_id,outcome.pred_spans,inst.gold,n_searches=outcome.n_searches,finished=outcome.finished)
   row.update({k:getattr(score,k) for k in ['file_recall','file_precision','file_f1','line_recall','found_all_files']})
   row.update(tools=list(teacher_manifest.tool_schema),repo=inst.repo,base_commit=inst.base_commit,language=inst.language,source=inst.source,events=outcome.events)
   validate_messages(row)
   try:weighted=sft.weight_record(row,inst.gold_files)
   except ValueError as exc:
    weighted=None;row["weighting_error"]=str(exc)
   # Match the historical Python release: only the system prompt/tool descriptions
   # are normalized to the exported student protocol, not assistant/tool content.
   # Raw/unfiltered ColGREP retains its strict teacher protocol, as in the source.
   student_messages=copy.deepcopy(row['messages']);student_messages[0]['content']=student_manifest.system_prompt
   if weighted:
    weighted['messages']=student_messages;weighted['tools']=list(student_manifest.tool_schema)
    assert len(weighted['turn_weights'])==sum(m['role']=='assistant' for m in row['messages'])
   row['_weighted']=weighted
   return row,None
  except Exception as e:return None,{'instance_id':inst.instance_id,'error':str(e)[:1500]}
 goal=min(target,4) if pilot else target
 errors=0
 with rawfile.open('a') as output,(state/'errors.jsonl').open('a') as err:
  for row,error in rolling_results(one,pending,max(0,goal-len(done)),min(workers,4) if pilot else workers):
   if error:err.write(json.dumps(error)+'\n');err.flush();errors+=1
   else:
    assert row['instance_id'] not in done
    output.write(json.dumps(row,ensure_ascii=False)+'\n');output.flush();done[row['instance_id']]=row
   status={'arm':arm,'new_unfiltered':len(done),'baseline_unfiltered':prep['baseline'][arm]['counts']['unfiltered'],'combined_unfiltered':len(done)+prep['baseline'][arm]['counts']['unfiltered'],'target':14500,'errors_this_run':errors,'pilot':pilot,'workers':workers}
   temporary=state/'progress.json.tmp';temporary.write_text(json.dumps(status,indent=2)+'\n');temporary.replace(state/'progress.json')
   print(json.dumps(status),flush=True)
 if len(done)<goal:raise RuntimeError(f'Candidate pool exhausted before {arm} target: {len(done)}/{goal}')
 if pilot:
  assert sum(r['n_searches']>0 for r in done.values())>=3,'Pilot search behavior failed'
  tool_errors=sum(bool(e['payload'].get('error')) for r in done.values() for e in r['events'] if e['kind']=='ToolCompleted')
  assert tool_errors<8,f'Pilot tool errors: {tool_errors}'
  if arm=='colgrep':assert any(r['retrieved_steps'] for r in done.values()),'Pilot produced no successful semantic retrieval'
 else:(state/'COMPLETE.json').write_text(json.dumps({'new_unfiltered':len(done),'combined_unfiltered':14500,'raw_sha256':hashlib.sha256(rawfile.read_bytes()).hexdigest()},indent=2)+'\n')
 llm._client.close();lock.close()

if __name__=='__main__':
 ap=argparse.ArgumentParser();ap.add_argument('--base-url',default='http://localhost:8001/v1');ap.add_argument('--preflight-only',action='store_true');ap.add_argument('--arm',choices=['grep','colgrep']);ap.add_argument('--workers',type=int,default=48);a=ap.parse_args()
 prep=json.loads((BASE/'preparation.json').read_text())
 for name,expected in prep['runtime_hashes'].items():assert hashlib.sha256((BASE/'runtime'/name).read_bytes()).hexdigest()==expected,name
 for arm in ['grep','colgrep']:
  teacher,student=manifests(arm);print(arm,'teacher prompt',teacher.prompt_hash,'student prompt',student.prompt_hash,flush=True)
 if a.preflight_only:raise SystemExit(0)
 vllm.assert_up(a.base_url)
 arms=[a.arm] if a.arm else ['grep','colgrep']
 for arm in arms:
  if not (BASE/arm/'progress.json').exists():run_arm(arm,a.base_url,pilot=True,workers=a.workers)
 with ThreadPoolExecutor(max_workers=len(arms)) as pool:
  list(pool.map(lambda arm:run_arm(arm,a.base_url,workers=a.workers),arms))
 print('Requested Python arms have 14,500 unfiltered trajectories',flush=True)