#!/usr/bin/env python3 """Resume-state helpers for SmartGridBench GCP fallback runs. The shell runner owns process orchestration. This helper owns the fiddly, testable parts of resume: trial classification, latency-row dedupe, atomic finalization, and manifest events. """ from __future__ import annotations import argparse import glob import hashlib import importlib.metadata import json import os import re import shlex import subprocess import sys from datetime import datetime, timezone from pathlib import Path from typing import Any _REPO_ROOT = Path(__file__).resolve().parent.parent if str(_REPO_ROOT) not in sys.path: sys.path.insert(0, str(_REPO_ROOT)) def _now_iso() -> str: return datetime.now(timezone.utc).isoformat() def _posix(path: Path | str) -> str: return Path(path).as_posix() def _norm(path: Path | str) -> str: return os.path.normpath(_posix(path)) def _shell_bool(value: bool) -> str: return "1" if value else "0" def _emit_shell(values: dict[str, Any]) -> None: for key, value in values.items(): print(f"{key}={shlex.quote(str(value))}") def _package_version(name: str) -> str | None: try: return importlib.metadata.version(name) except importlib.metadata.PackageNotFoundError: return None def _command_output(args: list[str]) -> str | None: try: return subprocess.check_output( args, stderr=subprocess.STDOUT, text=True, timeout=10, ).strip() except Exception: return None def collect_runtime_versions() -> dict[str, Any]: cuda_visible = os.environ.get("CUDA_VISIBLE_DEVICES") gpu_id = (cuda_visible or "0").split(",")[0].strip() or "0" versions: dict[str, Any] = { "vllm_version": _package_version("vllm"), "torch_version": _package_version("torch"), "cuda_visible_devices": cuda_visible, "nvidia_driver_version": None, "cuda_version": None, "nvidia_smi_query": None, } raw = _command_output( [ "nvidia-smi", f"--id={gpu_id}", "--query-gpu=driver_version,cuda_version", "--format=csv,noheader", ] ) if raw: versions["nvidia_smi_query"] = raw first_row = raw.splitlines()[0] parts = [part.strip() for part in first_row.split(",")] if parts: versions["nvidia_driver_version"] = parts[0] or None if len(parts) > 1: versions["cuda_version"] = parts[1] or None if versions["nvidia_driver_version"] is None: driver = _command_output( [ "nvidia-smi", f"--id={gpu_id}", "--query-gpu=driver_version", "--format=csv,noheader", ] ) if driver: versions["nvidia_driver_version"] = driver.splitlines()[0].strip() or None if versions["cuda_version"] is None: smi = _command_output(["nvidia-smi"]) if smi: match = re.search(r"CUDA Version:\s*([0-9.]+)", smi) if match: versions["cuda_version"] = match.group(1) return versions def _load_json_object(path: Path) -> tuple[dict[str, Any] | None, str | None]: if not path.exists(): return None, "missing_json" try: if path.stat().st_size == 0: return None, "zero_byte_json" payload = json.loads(path.read_text(encoding="utf-8")) except json.JSONDecodeError: return None, "invalid_json" except OSError as exc: return None, f"read_error:{exc}" if not isinstance(payload, dict): return None, "non_object_json" return payload, None def _scenario_payload(path: Path) -> dict[str, Any]: payload = json.loads(path.read_text(encoding="utf-8")) if not isinstance(payload, dict): raise ValueError(f"scenario is not a JSON object: {path}") return payload def _scenario_basename(path: Path) -> str: return path.stem def _expand_scenario_glob(scenario_glob: str) -> list[Path]: """Expand either one glob pattern or a shell-style list of scenario paths.""" tokens = shlex.split(scenario_glob) if not tokens: tokens = [scenario_glob] paths: list[Path] = [] seen: set[str] = set() for token in tokens: for match in sorted(glob.glob(token)): key = _norm(match) if key in seen: continue paths.append(Path(match)) seen.add(key) return paths def _step_failed(step: dict[str, Any]) -> bool: if step.get("success") is False: return True if step.get("error"): return True response = step.get("response") return isinstance(response, dict) and bool(response.get("error")) def derive_success(data: dict[str, Any]) -> bool | None: raw = data.get("success") if isinstance(raw, bool): return raw steps = data.get("history") or data.get("trajectory") or [] if not steps and not data.get("answer"): return None for step in steps: if isinstance(step, dict) and _step_failed(step): return False return bool(data.get("answer")) def _scenario_id_matches( payload: dict[str, Any], scenario: dict[str, Any], ) -> bool: embedded = payload.get("scenario") if not isinstance(embedded, dict): return True expected_id = scenario.get("id") actual_id = embedded.get("id") return not expected_id or not actual_id or expected_id == actual_id def _latency_records(latency_file: Path) -> list[dict[str, Any]]: if not latency_file.exists(): return [] records: list[dict[str, Any]] = [] for line in latency_file.read_text(encoding="utf-8").splitlines(): if not line.strip(): continue try: payload = json.loads(line) except json.JSONDecodeError: continue if isinstance(payload, dict): records.append(payload) return records def _record_matches_identity( record: dict[str, Any], scenario_file: Path, trial_index: int, output_path: Path, ) -> bool: if int(record.get("trial_index", -1)) != int(trial_index): return False output = record.get("output_path") if output and _norm(output) != _norm(output_path): return False scenario = record.get("scenario_file") return not scenario or _norm(scenario) == _norm(scenario_file) def _matching_latency_rows( records: list[dict[str, Any]], scenario_file: Path, trial_index: int, output_path: Path, ) -> list[dict[str, Any]]: return [ record for record in records if _record_matches_identity(record, scenario_file, trial_index, output_path) ] def _rows_conflict(rows: list[dict[str, Any]]) -> bool: if len(rows) < 2: return False normalized = { json.dumps(row, sort_keys=True, separators=(",", ":")) for row in rows } return len(normalized) > 1 def _candidate_trial_paths( run_dir: Path, scenario_file: Path, trial_index: int, output_path: Path, ) -> list[Path]: run_label = f"{_scenario_basename(scenario_file)}_run{trial_index:02d}.json" candidates = [output_path] candidates.extend(sorted(run_dir.glob(f"*_{run_label}"))) seen: set[str] = set() deduped: list[Path] = [] for candidate in candidates: key = str(candidate) if key not in seen: deduped.append(candidate) seen.add(key) return deduped def classify_trial( *, run_dir: Path, scenario_file: Path, trial_index: int, output_path: Path, latency_file: Path, require_latency: bool, ) -> dict[str, Any]: scenario = _scenario_payload(scenario_file) latency_records = _latency_records(latency_file) incomplete_reasons: list[str] = [] saw_candidate = False for candidate in _candidate_trial_paths( run_dir, scenario_file, trial_index, output_path ): if not candidate.exists(): continue saw_candidate = True payload, error = _load_json_object(candidate) if error is not None or payload is None: incomplete_reasons.append(error or "invalid_json") continue if not _scenario_id_matches(payload, scenario): incomplete_reasons.append("scenario_identity_conflict") continue success = derive_success(payload) if success is None: incomplete_reasons.append("success_not_derivable") continue rows = _matching_latency_rows( latency_records, scenario_file, trial_index, candidate ) if require_latency: if not rows: incomplete_reasons.append("missing_latency") continue if _rows_conflict(rows): incomplete_reasons.append("duplicate_conflicting_latency") continue return { "state": "complete_success" if success else "complete_failure", "complete": True, "success": success, "output_path": _posix(candidate), "reason": "terminal_trial", "latency_rows": len(rows), } stdout_candidates = [ path for path in run_dir.glob( f"*_{_scenario_basename(scenario_file)}_run{trial_index:02d}.json.stdout" ) if path.exists() ] if stdout_candidates and not saw_candidate: incomplete_reasons.append("dangling_stdout") return { "state": "incomplete", "complete": False, "success": False, "output_path": _posix(output_path), "reason": ",".join(sorted(set(incomplete_reasons))) or "missing_json", "latency_rows": 0, } def validate_run_artifacts( *, run_dir: Path, scenario_glob: str, trials: int, run_name: str = "", require_latency: bool = True, ) -> dict[str, Any]: scenario_files = _expand_scenario_glob(scenario_glob) expected = len(scenario_files) * int(trials) complete = 0 success = 0 failure = 0 missing: list[dict[str, Any]] = [] latency_file = run_dir / "latencies.jsonl" if expected == 0: return { "valid": False, "expected": 0, "complete": 0, "success": 0, "failure": 0, "missing": [], "reason": "no_expected_trials", } for scenario_file in scenario_files: basename = _scenario_basename(scenario_file) for trial_index in range(1, int(trials) + 1): output_name = f"{basename}_run{trial_index:02d}.json" if run_name: output_name = f"{run_name}_{output_name}" result = classify_trial( run_dir=run_dir, scenario_file=scenario_file, trial_index=trial_index, output_path=run_dir / output_name, latency_file=latency_file, require_latency=require_latency, ) if result["complete"]: complete += 1 if result["success"]: success += 1 else: failure += 1 continue missing.append( { "scenario_file": _posix(scenario_file), "trial_index": trial_index, "reason": result["reason"], "output_path": result["output_path"], } ) valid = complete == expected return { "valid": valid, "expected": expected, "complete": complete, "success": success, "failure": failure, "missing": missing, "reason": "ok" if valid else "incomplete_trajectory_artifacts", } def _sha256(path: Path) -> str | None: if not path.exists() or not path.is_file(): return None h = hashlib.sha256() with path.open("rb") as fh: for chunk in iter(lambda: fh.read(1024 * 1024), b""): h.update(chunk) return h.hexdigest() def write_manifest_event( *, manifest_file: Path, state: str, scenario_file: Path, trial_index: int, output_path: Path, run_name: str, reason: str = "", batch_id: str = "", latency_seconds: float | None = None, return_code: int | None = None, started_at: str = "", finished_at: str = "", extra: dict[str, Any] | None = None, ) -> None: manifest_file.parent.mkdir(parents=True, exist_ok=True) payload: dict[str, Any] = { "schema_version": 1, "batch_id": batch_id or None, "run_name": run_name, "event": "trial_resume_state", "state": state, "reason": reason or None, "scenario_file": _posix(scenario_file), "scenario_basename": _scenario_basename(scenario_file), "trial_index": int(trial_index), "output_path": _posix(output_path), "latency_seconds": latency_seconds, "return_code": return_code, "started_at": started_at or None, "finished_at": finished_at or _now_iso(), "compute_provider": os.environ.get("SMARTGRID_COMPUTE_PROVIDER"), "compute_zone": os.environ.get("SMARTGRID_COMPUTE_ZONE"), "compute_instance": os.environ.get("SMARTGRID_COMPUTE_INSTANCE"), "gpu_type": os.environ.get("GPU_TYPE"), "runtime_versions": collect_runtime_versions(), } if extra: payload.update(extra) with manifest_file.open("a", encoding="utf-8") as fh: fh.write(json.dumps(payload, sort_keys=True) + "\n") def preserve_incomplete_output(output_path: Path) -> Path | None: if not output_path.exists(): return None suffix = datetime.now(timezone.utc).strftime("%Y%m%dT%H%M%SZ") preserved = output_path.with_name(f"{output_path.name}.incomplete.{suffix}") output_path.replace(preserved) return preserved def _apply_final_answer_guards(payload: dict[str, Any]) -> None: try: from scripts.mitigation_guards import ( # type: ignore apply_explicit_fault_risk_adjudication, apply_missing_evidence_final_answer_guard, env_flag_enabled, ) except ImportError: return apply_missing_evidence_final_answer_guard( payload, enabled=env_flag_enabled(os.environ.get("ENABLE_MISSING_EVIDENCE_GUARD")), ) if "fault_risk_adjudication" not in payload: apply_explicit_fault_risk_adjudication( payload, enabled=env_flag_enabled( os.environ.get("ENABLE_EXPLICIT_FAULT_RISK_ADJUDICATION") ), ) def _upsert_latency_row( *, latency_file: Path, scenario_file: Path, trial_index: int, output_path: Path, latency_seconds: float, ) -> None: records = [ row for row in _latency_records(latency_file) if not _record_matches_identity(row, scenario_file, trial_index, output_path) ] records.append( { "scenario_file": _posix(scenario_file), "trial_index": int(trial_index), "latency_seconds": latency_seconds, "output_path": _posix(output_path), } ) latency_file.parent.mkdir(parents=True, exist_ok=True) tmp = latency_file.with_suffix(latency_file.suffix + ".tmp") tmp.write_text( "".join(json.dumps(record, sort_keys=True) + "\n" for record in records), encoding="utf-8", ) tmp.replace(latency_file) def finalize_trial( *, scenario_file: Path, trial_index: int, temp_output: Path, output_path: Path, latency_file: Path, manifest_file: Path, run_name: str, batch_id: str, start_epoch: float, end_epoch: float, return_code: int, ) -> dict[str, Any]: payload, error = _load_json_object(temp_output) if error is not None or payload is None: write_manifest_event( manifest_file=manifest_file, state="incomplete", scenario_file=scenario_file, trial_index=trial_index, output_path=output_path, run_name=run_name, reason=error or "invalid_json", batch_id=batch_id, return_code=return_code, ) return { "state": "incomplete", "success": False, "reason": error or "invalid_json", } scenario = _scenario_payload(scenario_file) payload["scenario"] = scenario _apply_final_answer_guards(payload) success = derive_success(payload) if success is None: write_manifest_event( manifest_file=manifest_file, state="incomplete", scenario_file=scenario_file, trial_index=trial_index, output_path=output_path, run_name=run_name, reason="success_not_derivable", batch_id=batch_id, return_code=return_code, ) return { "state": "incomplete", "success": False, "reason": "success_not_derivable", } payload["success"] = success output_path.parent.mkdir(parents=True, exist_ok=True) temp_output.write_text( json.dumps(payload, indent=2, default=str) + "\n", encoding="utf-8" ) original_hash = _sha256(output_path) temp_hash = _sha256(temp_output) temp_output.replace(output_path) divergent: bool | None if original_hash is None: divergent = None else: divergent = original_hash != temp_hash latency_seconds = float(end_epoch) - float(start_epoch) _upsert_latency_row( latency_file=latency_file, scenario_file=scenario_file, trial_index=trial_index, output_path=output_path, latency_seconds=latency_seconds, ) state = "complete_success" if success else "complete_failure" write_manifest_event( manifest_file=manifest_file, state=state, scenario_file=scenario_file, trial_index=trial_index, output_path=output_path, run_name=run_name, reason="trial_executed", batch_id=batch_id, latency_seconds=latency_seconds, return_code=return_code, extra={ "original_output_sha256": original_hash, "final_output_sha256": temp_hash, "divergent": divergent, }, ) return {"state": state, "success": success, "reason": "trial_executed"} def _build_parser() -> argparse.ArgumentParser: parser = argparse.ArgumentParser(description=__doc__) sub = parser.add_subparsers(dest="command", required=True) status = sub.add_parser("trial-status-shell") status.add_argument("--run-dir", required=True, type=Path) status.add_argument("--scenario-file", required=True, type=Path) status.add_argument("--trial-index", required=True, type=int) status.add_argument("--output-path", required=True, type=Path) status.add_argument("--latency-file", required=True, type=Path) status.add_argument("--require-latency", action="store_true") preserve = sub.add_parser("preserve-incomplete") preserve.add_argument("--output-path", required=True, type=Path) preserve.add_argument("--manifest-file", required=True, type=Path) preserve.add_argument("--scenario-file", required=True, type=Path) preserve.add_argument("--trial-index", required=True, type=int) preserve.add_argument("--run-name", required=True) preserve.add_argument("--batch-id", default="") preserve.add_argument("--reason", default="rerun_preserved_incomplete") finalize = sub.add_parser("finalize-trial-shell") finalize.add_argument("--scenario-file", required=True, type=Path) finalize.add_argument("--trial-index", required=True, type=int) finalize.add_argument("--temp-output", required=True, type=Path) finalize.add_argument("--output-path", required=True, type=Path) finalize.add_argument("--latency-file", required=True, type=Path) finalize.add_argument("--manifest-file", required=True, type=Path) finalize.add_argument("--run-name", required=True) finalize.add_argument("--batch-id", default="") finalize.add_argument("--start-epoch", required=True, type=float) finalize.add_argument("--end-epoch", required=True, type=float) finalize.add_argument("--return-code", required=True, type=int) event = sub.add_parser("manifest-event") event.add_argument("--manifest-file", required=True, type=Path) event.add_argument("--state", required=True) event.add_argument("--scenario-file", required=True, type=Path) event.add_argument("--trial-index", required=True, type=int) event.add_argument("--output-path", required=True, type=Path) event.add_argument("--run-name", required=True) event.add_argument("--reason", default="") event.add_argument("--batch-id", default="") validate = sub.add_parser("validate-run-shell") validate.add_argument("--run-dir", required=True, type=Path) validate.add_argument("--scenario-glob", required=True) validate.add_argument("--trials", required=True, type=int) validate.add_argument("--run-name", default="") validate.add_argument("--require-latency", action="store_true") return parser def main() -> None: args = _build_parser().parse_args() if args.command == "trial-status-shell": result = classify_trial( run_dir=args.run_dir, scenario_file=args.scenario_file, trial_index=args.trial_index, output_path=args.output_path, latency_file=args.latency_file, require_latency=args.require_latency, ) _emit_shell( { "RESUME_STATE": result["state"], "RESUME_COMPLETE": _shell_bool(bool(result["complete"])), "RESUME_SUCCESS": _shell_bool(bool(result["success"])), "RESUME_OUTPUT_PATH": result["output_path"], "RESUME_REASON": result["reason"], } ) return if args.command == "preserve-incomplete": preserved = preserve_incomplete_output(args.output_path) if preserved is not None: write_manifest_event( manifest_file=args.manifest_file, state="incomplete", scenario_file=args.scenario_file, trial_index=args.trial_index, output_path=preserved, run_name=args.run_name, reason=args.reason, batch_id=args.batch_id, ) return if args.command == "finalize-trial-shell": result = finalize_trial( scenario_file=args.scenario_file, trial_index=args.trial_index, temp_output=args.temp_output, output_path=args.output_path, latency_file=args.latency_file, manifest_file=args.manifest_file, run_name=args.run_name, batch_id=args.batch_id, start_epoch=args.start_epoch, end_epoch=args.end_epoch, return_code=args.return_code, ) _emit_shell( { "FINAL_STATE": result["state"], "FINAL_SUCCESS": _shell_bool(bool(result["success"])), "FINAL_REASON": result["reason"], } ) return if args.command == "manifest-event": write_manifest_event( manifest_file=args.manifest_file, state=args.state, scenario_file=args.scenario_file, trial_index=args.trial_index, output_path=args.output_path, run_name=args.run_name, reason=args.reason, batch_id=args.batch_id, ) return if args.command == "validate-run-shell": result = validate_run_artifacts( run_dir=args.run_dir, scenario_glob=args.scenario_glob, trials=args.trials, run_name=args.run_name, require_latency=args.require_latency, ) missing_sample = result["missing"][:5] _emit_shell( { "VALIDATION_PASSED": _shell_bool(bool(result["valid"])), "VALIDATION_EXPECTED": result["expected"], "VALIDATION_COMPLETE": result["complete"], "VALIDATION_SUCCESS": result["success"], "VALIDATION_FAILURE": result["failure"], "VALIDATION_MISSING": len(result["missing"]), "VALIDATION_REASON": result["reason"], "VALIDATION_MISSING_SAMPLE": json.dumps( missing_sample, sort_keys=True, separators=(",", ":") ), } ) if not result["valid"]: raise SystemExit(1) return raise AssertionError(args.command) if __name__ == "__main__": try: main() except Exception as exc: print(f"gcp_resume_state: ERROR: {exc}", file=sys.stderr) raise