337 lines
12 KiB
Python
337 lines
12 KiB
Python
#!/usr/bin/env python3
|
|
"""Replay the two Qwen235 Fixed-PD A1 cells with Frontier state outputs.
|
|
|
|
The frozen command is the experimental input. This runner changes only the
|
|
metrics output/run id and the two existing state-observation flags, then checks
|
|
that request-level scorer inputs are byte-identical to the frozen A1 run.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import argparse
|
|
import hashlib
|
|
import importlib.util
|
|
import json
|
|
import os
|
|
import subprocess
|
|
import sys
|
|
import time
|
|
from pathlib import Path
|
|
from typing import Any
|
|
|
|
|
|
HERE = Path(__file__).resolve().parent
|
|
REPO_ROOT = HERE.parents[1]
|
|
TELEMETRY_ROOT = REPO_ROOT / "runs/telemetry-residual"
|
|
sys.path.insert(0, str(TELEMETRY_ROOT))
|
|
|
|
from common_state import summarize_frontier # noqa: E402
|
|
from run_frontier_state import enable_state_outputs, find_state_metrics # noqa: E402
|
|
|
|
|
|
CONFIGS = ("tp4_ep1_mns64", "tp8_ep8_mns64")
|
|
ALLOWED_FLAG_CHANGES = {
|
|
"--metrics_config_output_dir",
|
|
"--metrics_config_run_id",
|
|
"--no-metrics_config_store_frontier_stage_batch_ledger",
|
|
"--metrics_config_store_frontier_stage_batch_ledger",
|
|
"--no-metrics_config_keep_individual_batch_metrics",
|
|
"--metrics_config_keep_individual_batch_metrics",
|
|
}
|
|
|
|
|
|
def load_module(name: str, path: Path):
|
|
spec = importlib.util.spec_from_file_location(name, path)
|
|
if spec is None or spec.loader is None:
|
|
raise ImportError(path)
|
|
module = importlib.util.module_from_spec(spec)
|
|
sys.modules[spec.name] = module
|
|
spec.loader.exec_module(module)
|
|
return module
|
|
|
|
|
|
def sha256_file(path: Path) -> str:
|
|
digest = hashlib.sha256()
|
|
with path.open("rb") as source:
|
|
for chunk in iter(lambda: source.read(1 << 20), b""):
|
|
digest.update(chunk)
|
|
return digest.hexdigest()
|
|
|
|
|
|
def atomic_json(path: Path, payload: Any) -> None:
|
|
path.parent.mkdir(parents=True, exist_ok=True)
|
|
temporary = path.with_suffix(path.suffix + ".tmp")
|
|
temporary.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
|
|
os.replace(temporary, path)
|
|
|
|
|
|
def replace_option(command: list[str], option: str, value: str) -> list[str]:
|
|
result = list(command)
|
|
if result.count(option) != 1:
|
|
raise ValueError(f"expected exactly one {option}, found {result.count(option)}")
|
|
index = result.index(option)
|
|
if index + 1 >= len(result) or result[index + 1].startswith("--"):
|
|
raise ValueError(f"{option} has no value")
|
|
result[index + 1] = value
|
|
return result
|
|
|
|
|
|
def option_value(command: list[str], option: str) -> str:
|
|
if command.count(option) != 1:
|
|
raise ValueError(f"expected exactly one {option}, found {command.count(option)}")
|
|
return command[command.index(option) + 1]
|
|
|
|
|
|
def _semantic_options(command: list[str]) -> dict[str, tuple[str, ...]]:
|
|
"""Parse CLI tokens sufficiently to audit this controlled command edit."""
|
|
|
|
result: dict[str, tuple[str, ...]] = {}
|
|
index = 0
|
|
while index < len(command):
|
|
token = command[index]
|
|
if not token.startswith("--"):
|
|
index += 1
|
|
continue
|
|
values: list[str] = []
|
|
index += 1
|
|
while index < len(command) and not command[index].startswith("--"):
|
|
values.append(command[index])
|
|
index += 1
|
|
if token in result:
|
|
raise ValueError(f"duplicate option in frozen command: {token}")
|
|
result[token] = tuple(values)
|
|
return result
|
|
|
|
|
|
def transform_command(
|
|
command: list[str], *, metrics_root: Path, run_id: str
|
|
) -> list[str]:
|
|
result = replace_option(
|
|
command, "--metrics_config_output_dir", str(metrics_root.resolve())
|
|
)
|
|
result = replace_option(result, "--metrics_config_run_id", run_id)
|
|
result = enable_state_outputs(result)
|
|
|
|
before = _semantic_options(command)
|
|
after = _semantic_options(result)
|
|
changed = {
|
|
option
|
|
for option in set(before) | set(after)
|
|
if before.get(option) != after.get(option)
|
|
}
|
|
if not changed <= ALLOWED_FLAG_CHANGES:
|
|
raise ValueError(f"unapproved command changes: {sorted(changed - ALLOWED_FLAG_CHANGES)}")
|
|
required = {
|
|
"--metrics_config_output_dir",
|
|
"--metrics_config_run_id",
|
|
"--no-metrics_config_store_frontier_stage_batch_ledger",
|
|
"--metrics_config_store_frontier_stage_batch_ledger",
|
|
"--metrics_config_keep_individual_batch_metrics",
|
|
}
|
|
if not required <= changed:
|
|
raise ValueError(f"required controlled changes missing: {sorted(required - changed)}")
|
|
return result
|
|
|
|
|
|
def environment(frontier_source: Path, python_deps: Path) -> dict[str, str]:
|
|
result = os.environ.copy()
|
|
result.update(
|
|
{
|
|
"PYTHONPATH": ":".join([str(python_deps), str(frontier_source)]),
|
|
"CUDA_VISIBLE_DEVICES": "",
|
|
"NVIDIA_VISIBLE_DEVICES": "void",
|
|
"WANDB_DISABLED": "true",
|
|
"VIDUR_DISABLE_WANDB": "1",
|
|
"FRONTIER_LOG_LEVEL": "WARNING",
|
|
"PYTHONDONTWRITEBYTECODE": "1",
|
|
}
|
|
)
|
|
return result
|
|
|
|
|
|
def parse_args() -> argparse.Namespace:
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--base-sim-root", type=Path, required=True)
|
|
parser.add_argument("--output-root", type=Path, required=True)
|
|
parser.add_argument("--frontier-source", type=Path, required=True)
|
|
parser.add_argument("--python-deps", type=Path, required=True)
|
|
parser.add_argument("--config", action="append", choices=CONFIGS)
|
|
parser.add_argument("--timeout-seconds", type=float, default=1200)
|
|
parser.add_argument("--resume", action="store_true")
|
|
return parser.parse_args()
|
|
|
|
|
|
def run_cell(
|
|
*,
|
|
config: str,
|
|
base_sim_root: Path,
|
|
output_root: Path,
|
|
frontier_source: Path,
|
|
python_deps: Path,
|
|
timeout_seconds: float,
|
|
resume: bool,
|
|
q30: Any,
|
|
) -> dict[str, Any]:
|
|
base_run = base_sim_root / "runs" / config / "eval"
|
|
base_command_path = base_run / "command.json"
|
|
base_result_path = base_run / "result.json"
|
|
if not base_command_path.is_file() or not base_result_path.is_file():
|
|
raise FileNotFoundError(f"frozen A1 inputs missing under {base_run}")
|
|
base_command = json.loads(base_command_path.read_text())
|
|
base_result = json.loads(base_result_path.read_text())
|
|
if base_result.get("status") != "completed":
|
|
raise ValueError(f"frozen A1 result is not completed: {base_result_path}")
|
|
|
|
run_root = output_root / config
|
|
result_path = run_root / "result.json"
|
|
if resume and result_path.is_file():
|
|
previous = json.loads(result_path.read_text())
|
|
if (
|
|
previous.get("status") == "PASS"
|
|
and previous.get("inputs", {}).get("base_command_sha256")
|
|
== sha256_file(base_command_path)
|
|
and previous.get("inputs", {}).get("base_result_sha256")
|
|
== sha256_file(base_result_path)
|
|
):
|
|
return previous
|
|
if run_root.exists() and any(run_root.iterdir()):
|
|
raise FileExistsError(f"refusing non-empty output: {run_root}")
|
|
run_root.mkdir(parents=True, exist_ok=True)
|
|
|
|
command = transform_command(
|
|
base_command,
|
|
metrics_root=run_root / "frontier_metrics",
|
|
run_id=f"qwen235_fixed_pd_state_{config}",
|
|
)
|
|
atomic_json(run_root / "command.json", command)
|
|
manifest = {
|
|
"schema": "qwen235-fixed-pd-state-replay-v1",
|
|
"config": config,
|
|
"inputs": {
|
|
"base_run": str(base_run.resolve()),
|
|
"base_command_sha256": sha256_file(base_command_path),
|
|
"base_result_sha256": sha256_file(base_result_path),
|
|
"base_request_metrics_sha256": base_result["request_metrics_sha256"],
|
|
},
|
|
"controlled_changes": [
|
|
"metrics output directory",
|
|
"metrics run id",
|
|
"full Frontier stage/batch ledger enabled",
|
|
"individual batch metrics enabled",
|
|
],
|
|
"command_sha256": sha256_file(run_root / "command.json"),
|
|
"frontier": {
|
|
"source": str(frontier_source),
|
|
"git_head": subprocess.check_output(
|
|
["git", "-C", str(frontier_source), "rev-parse", "HEAD"], text=True
|
|
).strip(),
|
|
},
|
|
"environment": {
|
|
"PYTHONPATH": ":".join([str(python_deps), str(frontier_source)]),
|
|
"CUDA_VISIBLE_DEVICES": "",
|
|
"NVIDIA_VISIBLE_DEVICES": "void",
|
|
"FRONTIER_LOG_LEVEL": "WARNING",
|
|
},
|
|
}
|
|
atomic_json(run_root / "run_manifest.json", manifest)
|
|
|
|
started = time.monotonic()
|
|
with (run_root / "stdout.log").open("w") as stdout, (
|
|
run_root / "stderr.log"
|
|
).open("w") as stderr:
|
|
try:
|
|
completed = subprocess.run(
|
|
command,
|
|
cwd=frontier_source,
|
|
env=environment(frontier_source, python_deps),
|
|
stdout=stdout,
|
|
stderr=stderr,
|
|
timeout=timeout_seconds,
|
|
check=False,
|
|
)
|
|
returncode = int(completed.returncode)
|
|
except subprocess.TimeoutExpired:
|
|
returncode = 124
|
|
elapsed_seconds = time.monotonic() - started
|
|
if returncode != 0:
|
|
failure = {
|
|
"status": "STOP",
|
|
"config": config,
|
|
"returncode": returncode,
|
|
"elapsed_seconds": elapsed_seconds,
|
|
}
|
|
atomic_json(run_root / "failure.json", failure)
|
|
raise RuntimeError(f"state replay failed: {failure}")
|
|
|
|
fallback_evidence = q30.collective_fallback_evidence(run_root)
|
|
if fallback_evidence:
|
|
raise RuntimeError(f"collective-profile fallback detected: {fallback_evidence}")
|
|
paths = find_state_metrics(run_root)
|
|
request_metrics_sha256 = sha256_file(paths["requests"])
|
|
scorer_equivalent = request_metrics_sha256 == base_result["request_metrics_sha256"]
|
|
if not scorer_equivalent:
|
|
raise RuntimeError(
|
|
f"observation changed scorer input for {config}: "
|
|
f"{request_metrics_sha256} != {base_result['request_metrics_sha256']}"
|
|
)
|
|
state = summarize_frontier(
|
|
system_metrics_path=paths["system"],
|
|
request_metrics_path=paths["requests"],
|
|
batch_metrics_path=paths["batches"],
|
|
ledger_path=paths["ledger"],
|
|
)
|
|
atomic_json(run_root / "common-state.json", state)
|
|
result = {
|
|
"schema": "qwen235-fixed-pd-state-replay-result-v1",
|
|
"status": "PASS",
|
|
"config": config,
|
|
"elapsed_seconds": elapsed_seconds,
|
|
"returncode": returncode,
|
|
"inputs": manifest["inputs"],
|
|
"scorer_equivalence": {
|
|
"request_metrics_byte_identical": scorer_equivalent,
|
|
"request_metrics_sha256": request_metrics_sha256,
|
|
"base_metrics": base_result["metrics"],
|
|
},
|
|
"state_artifacts": {
|
|
name: {"path": str(path.resolve()), "sha256": sha256_file(path)}
|
|
for name, path in paths.items()
|
|
},
|
|
"common_state": state,
|
|
"collective_fallback_evidence": fallback_evidence,
|
|
}
|
|
atomic_json(result_path, result)
|
|
return result
|
|
|
|
|
|
def main() -> None:
|
|
args = parse_args()
|
|
for name in ("base_sim_root", "output_root", "frontier_source", "python_deps"):
|
|
setattr(args, name, getattr(args, name).resolve())
|
|
selected = tuple(args.config or CONFIGS)
|
|
q30 = load_module("q235_state_q30", HERE / "run_frontier_qwen30_exact_trace_surface.py")
|
|
results = []
|
|
for config in selected:
|
|
result = run_cell(
|
|
config=config,
|
|
base_sim_root=args.base_sim_root,
|
|
output_root=args.output_root,
|
|
frontier_source=args.frontier_source,
|
|
python_deps=args.python_deps,
|
|
timeout_seconds=args.timeout_seconds,
|
|
resume=args.resume,
|
|
q30=q30,
|
|
)
|
|
results.append(result)
|
|
print(json.dumps({"config": config, "status": result["status"]}), flush=True)
|
|
aggregate = {
|
|
"schema": "qwen235-fixed-pd-state-replay-aggregate-v1",
|
|
"status": "PASS" if all(result["status"] == "PASS" for result in results) else "STOP",
|
|
"results": results,
|
|
}
|
|
atomic_json(args.output_root / "state_replay.json", aggregate)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|