data-gen-bpu/scripts/sync_status.py

150 lines
7.1 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#!/usr/bin/env python3
"""sync_status —— 把 registry/models.csv 的 status 列刷新成实机事实。
PYTHONPATH=. python3 scripts/sync_status.py # dry-run:只打印将发生的变化
PYTHONPATH=. python3 scripts/sync_status.py --apply # 真正写 models.csv
models.csv 的 status 是 Excel 导入时(excel_to_models.py)按来源启发式打的:
torchvision/cls→ready、ultralytics/det→ready、其余 needs_adapter/needs_eval/skip。
跑了一阵子之后它和实机进度脱节:适配器、probe、t1/calibration 都陆续补齐了,
但 status 没人回头改,于是 `bench expand` 默认只展开 status=ready,看不到这些
实际已具备准入条件的模型。
判定不重新发明:直接复用 harness/coverage.py 的 model_blockers()(与
master_inventory 同一调用方式:models() → model_spec() → model_blockers(row, cache))。
规则:
blockers 为空 且 status ∈ {needs_adapter, needs_eval} → 改 ready
blockers 非空 → 保持原 status
status = skip → 一律不动(非神经网络,不量化)
status = ready → 不动(只升不降;是否有 blocker
与 label 无关,不做降级)
写入用 csv 模块 round-trip:列结构、行序、其他列字节不变。写入前先验证未修改
内容能被同一套 writer 逐字节还原(防格式漂移),再只改 status 单元格,
临时文件 + os.replace 原子落盘。
默认 dry-run 的原因:models.csv 参与 registry_fingerprint(),跑着的 collector /
scheduler 对指纹变化敏感——当前 run 有 rollout 在采时改 registry 可能扰动它们。
何时可以 --apply:当前 run 收尾(rollout 全部终态、调度进程退出)之后再跑;
改完后用 `PYTHONPATH=. python3 -c "from harness.registry import status_counts; print(status_counts())"`
复核,再正常 scripts/start_run.sh --resume。
"""
from __future__ import annotations
import argparse
import csv
import io
import os
import sys
import tempfile
import time
from collections import Counter
from pathlib import Path
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from harness import paths # noqa: E402
from harness.adapters import model_spec # noqa: E402
from harness.coverage import model_blockers # noqa: E402
from harness.registry import models, status_counts # noqa: E402
PROMOTE = ("needs_adapter", "needs_eval")
def roundtrip(rows: list[list[str]], terminator: str) -> bytes:
"""csv 模块序列化;与读入同一 dialect,只有 terminator 按文件实测。"""
buf = io.StringIO(newline="")
csv.writer(buf, lineterminator=terminator).writerows(rows)
return buf.getvalue().encode("utf-8")
def main() -> int:
ap = argparse.ArgumentParser(description="按 model_blockers() 实况刷新 models.csv 的 status")
ap.add_argument("--csv", default=str(paths.REGISTRY / "models.csv"),
help="models.csv 路径(默认 registry/models.csv)")
ap.add_argument("--apply", action="store_true",
help="真正写入(默认 dry-run 只打印)")
a = ap.parse_args()
csv_path = Path(a.csv)
raw = csv_path.read_bytes()
terminator = "\r\n" if b"\r\n" in raw else "\n"
text = raw.decode("utf-8")
rows = list(csv.reader(io.StringIO(text, newline="")))
pristine = [list(r) for r in rows] # 改动前的逐字节快照
header = rows[0]
col = {name: i for i, name in enumerate(header)}
for need in ("model_id", "task", "status"):
if need not in col:
raise SystemExit(f"{csv_path} 缺列 {need!r},现有:{header}")
# ---- 判定:与 master_inventory() 相同的调用链,共享一个 prerequisite cache
cache: dict = {}
t0 = time.time()
admission: dict[str, list[str]] = {}
for row in models():
spec = model_spec(row)
admission[spec["model_id"]] = model_blockers(spec, cache)
elapsed = time.time() - t0
# ---- 规则应用(改的是原始 CSV 行,不是 models() 的 strip 副本)
st_before = Counter((r[col["status"]].strip() or "ready")
for r in pristine[1:] if r[col["model_id"]].strip())
changes: list[tuple[str, str, str, str]] = [] # (model, task, old, new)
for r in rows[1:]:
mid = r[col["model_id"]].strip()
if not mid or mid not in admission:
continue
old = r[col["status"]].strip()
if old in ("skip", "ready"):
continue # skip 不动;ready 只升不降
if not admission[mid] and old in PROMOTE:
r[col["status"]] = "ready"
changes.append((mid, r[col["task"]].strip(), old, "ready"))
# ---- 报告
print(f"models.csv: {csv_path} · {len(rows) - 1} 行 · model_blockers() 实测 {len(admission)} 个模型,耗时 {elapsed:.1f}s")
print(f"status 现状: {dict(st_before.most_common())}")
print()
if not changes:
print("没有需要变更的模型(blockers 为空的都已 ready,其余按规则保持)")
return 0
for mid, task, old, new in changes:
print(f" {mid:<42} {task:<6} {old:<14} → {new} 依据: model_blockers()=0(适配器/probe/t1/calibration 就绪)")
st_after = dict(st_before)
for _, _, old, new in changes:
st_after[old] -= 1
st_after[new] += 1
print()
print(f"将变更 {len(changes)} 个模型 · 按 task: {dict(Counter(t for _, t, _, _ in changes))}")
print(f"变更后 status: {dict(Counter({k: v for k, v in st_after.items() if v}).most_common())}")
if not a.apply:
print("\ndry-run,未写盘。确认后加 --apply(等当前 run 收尾、rollout 全部终态之后再动 registry)。")
return 0
# ---- 写入:先证明这套 writer 配置能逐字节还原原文件,再落盘
if roundtrip(pristine, terminator) != raw:
raise SystemExit("round-trip 校验失败:csv 序列化无法逐字节还原原文件,拒绝写入(防格式漂移)")
fd, tmp = tempfile.mkstemp(dir=csv_path.parent, prefix=csv_path.name + ".")
try:
with os.fdopen(fd, "w", newline="", encoding="utf-8") as stream:
csv.writer(stream, lineterminator=terminator).writerows(rows)
os.replace(tmp, csv_path)
finally:
if os.path.exists(tmp):
os.unlink(tmp)
print(f"\nwrote {len(changes)} 处 status 变更 → {csv_path}")
with csv_path.open(encoding="utf-8", newline="") as stream:
written = list(csv.reader(stream))
if written != rows:
raise SystemExit("写入后回读与预期行不一致,请立即检查文件")
if csv_path == paths.REGISTRY / "models.csv":
models.cache_clear()
print(f"复核 registry.models() 重新加载: {status_counts()}") # 只认 REGISTRY 下的真身
else:
print(f"复核回读: {dict(Counter((r[col['status']].strip() or 'ready') for r in written[1:] if r[col['model_id']].strip()).most_common())}")
return 0
if __name__ == "__main__":
sys.exit(main())