#!/usr/bin/env python3 """fai_bench 题包自检 —— 只读,任何失败退出码非 0。 完全自洽:所有路径从本脚本位置推导,不依赖任何外部台账或源树, 所以在你 clone 下来的任何位置都能直接跑。 python3 scripts/verify_package.py # 全量 python3 scripts/verify_package.py kfc lh # 只查某些子集 检查项(逐题): 1 必备件齐全 instruction.md · task.toml · environment/Dockerfile · tests/ 2 task.toml 可解析 且可编辑范围 / 资源声明在 3 判分面可用 tests/test.sh 存在;性能题的锚点能被解析到(不会静默回落 1.0) 4 Dockerfile 可解析 复刻 Dockerfile 解析器:heredoc 配对、无悬挂续行、 COPY 源都在构建上下文里且没被 .dockerignore 排除, 每个 RUN 的 shell 体过 `bash -n` 5 锚点自洽 tests/ref_speedup.txt(若有)与 caveat 记录的值一致 6 挖洞树自证 有 oracle.patch 的题:补丁必须能正向打在 environment/repo/ 上、反向打不上 7 卫生 无 __pycache__ / *.pyc / *.bak* / .DS_Store / 编辑器残留 8 可运行三件套 pre_artifacts.sh 存在且 `bash -n` 过;solution/solve.sh 存在、可执行、 四态 CLI(--negative/--baseline2/--noop)且 `bash -n` 过; task.toml 是 schema_version="2.0" """ from __future__ import annotations import glob import os import re import subprocess import sys import tempfile try: import tomllib as toml_r # py >= 3.11 except ModuleNotFoundError: try: import tomli as toml_r # type: ignore except ModuleNotFoundError: toml_r = None # stdlib-only: schema check degrades def _toml_load(path): """Parse task.toml. Prefer tomllib/tomli; fall back to a scan for the handful of keys this checker needs so the checker itself stays dependency-free on py<3.11.""" if toml_r is not None: with open(path, "rb") as fh: return toml_r.load(fh) txt = open(path, "r", encoding="utf-8", errors="ignore").read() meta: dict = {} m = re.search(r'^\s*schema_version\s*=\s*["\']([^"\']+)', txt, re.M) if m: meta["schema_version"] = m.group(1) return meta ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) TASKS = os.path.join(ROOT, "tasks") BENCHES = [b for b in ("kfc", "lh", "e2e") if os.path.isdir(os.path.join(TASKS, b))] WANT = [a for a in sys.argv[1:] if not a.startswith("-")] or BENCHES KW = ("FROM RUN CMD LABEL MAINTAINER EXPOSE ENV ADD COPY ENTRYPOINT VOLUME USER " "WORKDIR ARG ONBUILD STOPSIGNAL HEALTHCHECK SHELL").split() KWRE = re.compile(r'^\s*(' + "|".join(KW) + r')(\s|$)', re.I) HDRE = re.compile(r'<<(-?)\s*(["\']?)([A-Za-z_][A-Za-z0-9_]*)\2') JUNK = re.compile(r'(^|/)__pycache__(/|$)|\.pyc$|\.pyo$|\.bak(_|\.|$)|\.orig(_|$)' r'|\.DS_Store$|~$|\.swp$|\.pre_') fails: list[str] = [] warns: list[str] = [] def fail(pkg: str, msg: str) -> None: fails.append(f"{pkg}: {msg}") def warn(pkg: str, msg: str) -> None: warns.append(f"{pkg}: {msg}") # ---------------------------------------------------------------- .dockerignore def _pat2re(pp: str) -> re.Pattern: """Docker 的 .dockerignore 语义:相对上下文根整路径匹配(支持 ** / * / ?)。""" out, i = "", 0 while i < len(pp): if pp.startswith("**", i): out += ".*" i += 2 if i < len(pp) and pp[i] == "/": out += "/?" i += 1 elif pp[i] == "*": out += "[^/]*" i += 1 elif pp[i] == "?": out += "[^/]" i += 1 else: out += re.escape(pp[i]) i += 1 return re.compile(r'^' + out + r'$') def dockerignored(rel: str, pats: list[str]) -> bool: rel = rel.strip("/") hit = False for p in pats: neg = p.startswith("!") rx = _pat2re((p[1:] if neg else p).strip("/")) m = bool(rx.match(rel)) if not m: # 命中父目录 -> 内容也被排除 parts = rel.split("/") m = any(rx.match("/".join(parts[:k])) for k in range(1, len(parts))) if m: hit = not neg return hit # ---------------------------------------------------------------- Dockerfile def check_dockerfile(pkg: str, df: str) -> None: ctx = os.path.dirname(os.path.dirname(df)) # 上下文 = 题包根 lines = open(df, errors="ignore").read().split("\n") di = os.path.join(ctx, ".dockerignore") pats = [l.strip() for l in open(di, errors="ignore")] if os.path.isfile(di) else [] pats = [p for p in pats if p and not p.startswith("#")] rel_df = os.path.relpath(df, ROOT) i, n = 0, len(lines) while i < n: raw = lines[i] if not raw.strip() or raw.lstrip().startswith("#"): i += 1 continue if not KWRE.match(raw): fail(pkg, f"{rel_df}:{i+1} 非指令行出现在指令位置(Dockerfile 解析会失败): " f"«{raw.strip()[:70]}»") i += 1 continue instr = KWRE.match(raw).group(1).upper() cont = [raw] while cont[-1].rstrip().endswith("\\"): if i + 1 >= n: fail(pkg, f"{rel_df}:{i+1} 文件末尾悬挂 `\\` 续行") break i += 1 cont.append(lines[i]) joined = "\n".join(cont) i += 1 hds: list[tuple[str, list[str]]] = [] for m in HDRE.finditer(joined): term, dash, body, ok = m.group(3), m.group(1) == "-", [], False while i < n: s = lines[i] if (s.lstrip("\t") if dash else s).rstrip("\r") == term: ok = True i += 1 break body.append(s) i += 1 if not ok: fail(pkg, f"{rel_df} heredoc `{term}` 未终止") hds.append((term, body)) if instr == "RUN": first = re.sub(r'^\s*RUN\s+', '', joined, flags=re.I) if not first.lstrip().startswith("["): # exec-form 不是 shell if hds and re.match(r'^<<-?\s*["\']?[A-Za-z_]', first.strip()): body = "\n".join(hds[0][1]) # RUN <<'SH' 脚本形式 else: body = re.sub(r'\\\s*\n', ' ', first) # shell 语义:\+换行 整对删除 for term, b in hds: body += "\n" + "\n".join(b) + "\n" + term for k, l in enumerate(joined.split("\n")): if k and re.search(r'(^|\s)#', l) and l.rstrip().endswith("\\"): fail(pkg, f"{rel_df} `\\`-续行的 RUN 里有内联 `#`,会吃掉后续命令") break with tempfile.NamedTemporaryFile("w", suffix=".sh", delete=False) as f: f.write(body) tmp = f.name r = subprocess.run(["bash", "-n", tmp], capture_output=True, text=True) os.unlink(tmp) if r.returncode: tail = (r.stderr.strip().split("\n") or [""])[-1][:90] fail(pkg, f"{rel_df} RUN 的 shell 体语法错: {tail}") if instr in ("COPY", "ADD") and "--from=" not in joined: toks = [t for t in re.sub(r'^\s*(COPY|ADD)\s+', '', joined, flags=re.I) .replace("\\\n", " ").replace("\n", " ").split() if not t.startswith("--")] for src in toks[:-1]: if re.match(r'^https?://', src): continue s = src.rstrip("/") if not glob.glob(os.path.join(ctx, s)): fail(pkg, f"{rel_df} COPY 源不在构建上下文里: {src}") elif dockerignored(s, pats): fail(pkg, f"{rel_df} COPY 源被 .dockerignore 排除: {src}") # ---------------------------------------------------------------- runnable three-piece def _bash_n(path: str) -> str: """Return '' if `bash -n ` is clean, else the last stderr line.""" r = subprocess.run(["bash", "-n", path], capture_output=True, text=True) if r.returncode == 0: return "" return (r.stderr.strip().split("\n") or [""])[-1][:90] or f"rc={r.returncode}" def check_runnable(pkg: str, root: str, meta: dict) -> None: """Gate 8: the runnable contract layered on top of the read/build package — pre_artifacts.sh, solution/solve.sh (four-mode CLI), and schema_version 2.0.""" pre = os.path.join(root, "pre_artifacts.sh") if not os.path.isfile(pre): fail(pkg, "缺 pre_artifacts.sh(提交契约:把工作树捕获成 model.patch)") else: err = _bash_n(pre) if err: fail(pkg, f"pre_artifacts.sh 语法错: {err}") solve = os.path.join(root, "solution", "solve.sh") if not os.path.isfile(solve): fail(pkg, "缺 solution/solve.sh(oracle 一键落地脚本)") else: if not os.access(solve, os.X_OK): fail(pkg, "solution/solve.sh 没有可执行位") err = _bash_n(solve) if err: fail(pkg, f"solution/solve.sh 语法错: {err}") body = open(solve, errors="ignore").read() # a solve.sh that only prints "no reference implementation" and exits 2 is a # legitimate shape for the 2 tasks that ship no oracle -- accept it as-is. if re.search(r'exit\s+2\b', body) and "--negative" not in body: pass else: for flag in ("--negative", "--baseline2", "--noop"): if flag not in body: warn(pkg, f"solution/solve.sh 未处理 {flag} 模式(四态 CLI 不全)") sv = str(meta.get("schema_version") or "") if sv != "2.0": warn(pkg, f"task.toml schema_version={sv or '缺失'}(期望 '2.0')") # ---------------------------------------------------------------- 每题 def check_pkg(bench: str, task: str) -> None: base = os.path.join(TASKS, bench, task) pkg = f"{bench}/{task}" root = base if os.path.isdir(os.path.join(base, "tests")) else os.path.join(base, "task") meta: dict = {} for rel in ("instruction.md", "task.toml", "environment", "tests"): if not os.path.exists(os.path.join(root, rel)): fail(pkg, f"缺 {rel}") dfs = sorted(glob.glob(os.path.join(root, "environment", "Dockerfile*"))) if not dfs: fail(pkg, "缺 environment/Dockerfile") tt = os.path.join(root, "task.toml") if os.path.isfile(tt): try: meta = _toml_load(tt) except Exception as e: fail(pkg, f"task.toml 解析失败: {e}") meta = {} D = lambda v: v if isinstance(v, dict) else {} gpus = D(meta.get("environment")).get("gpus", D(meta.get("resources")).get("gpus", meta.get("gpus"))) if gpus is None: warn(pkg, "task.toml 没有声明 gpus(检查过 [environment] / [resources] / 顶层)") vr = D(meta.get("verifier")) scope = (D(vr.get("entry")).get("primary_edit_paths") or vr.get("primary_edit_paths") or meta.get("primary_edit_paths")) if not scope: # 部分题的可编辑范围只在 tests/test.sh 里声明 tsp = os.path.join(root, "tests", "test.sh") if os.path.isfile(tsp): body = open(tsp, errors="ignore").read() if re.search(r'^\s*SCOPE=\(', body, re.M) or re.search(r'^\s*SUB=', body, re.M): scope = [""] if not scope: frozen = any(os.path.isfile(os.path.join(root, x)) for x in ("tests/.frozen_hashes.json", "tests/verifier-correctness-manifest.json", "tests/reward_manifest.json")) if not frozen: warn(pkg, "可编辑范围无处声明,且没有 sha256 冻结面 —— 无从判断哪些文件可改") # 有冻结面 = Type-3 端到端题的正常形制(整树可改,只冻结评分面),不告警 ts = os.path.join(root, "tests", "test.sh") if not os.path.isfile(ts): fail(pkg, "缺 tests/test.sh") else: s = open(ts, errors="ignore").read() if "ref_speedup" in s: got = None rt = os.path.join(root, "tests", "ref_speedup.txt") if os.path.isfile(rt): try: got = float(re.sub(r'[^0-9.]', '', open(rt).read()) or 0) except ValueError: got = None if got is None: for mf in ("tests/verifier-correctness-manifest.json", "environment/verifier-correctness-manifest.json", "tests/reward_manifest.json"): p = os.path.join(root, mf) if os.path.isfile(p): m = re.search(r'"ref_speedup"\s*:\s*(?:\{[^}]*"value"\s*:\s*)?([0-9.]+)', open(p, errors="ignore").read()) if m and float(m.group(1)) > 1: got = float(m.group(1)) break hard = re.search(r'ref_speedup=([0-9]+\.[0-9]+)', s) if got is None and not hard: fail(pkg, "性能题但锚点无处可解析(会静默回落 1.0 → 所有 reward 归 0)") elif got is not None and got <= 1: fail(pkg, f"锚点 <= 1({got}) → reward 恒为 0") for df in dfs: check_dockerfile(pkg, df) rt = os.path.join(root, "tests", "ref_speedup.txt") cv = os.path.join(root, "tests", "ref_speedup.caveat.md") if os.path.isfile(rt) and os.path.isfile(cv): v = re.sub(r'[^0-9.]', '', open(rt).read()) if v and v not in open(cv, errors="ignore").read(): warn(pkg, f"caveat.md 里没出现 ref_speedup.txt 的值 {v}(两处可能已脱钩)") repo = os.path.join(root, "environment", "repo") for cand in ("solution/oracle.patch", "oracle.patch", "tests/oracle.patch"): op = os.path.join(root, cand) if os.path.isfile(op) and os.path.isdir(repo): fwd = subprocess.run(["git", "apply", "--check", "-p1", op], cwd=repo, capture_output=True).returncode == 0 rev = subprocess.run(["git", "apply", "--check", "-R", "-p1", op], cwd=repo, capture_output=True).returncode == 0 if not fwd: fail(pkg, f"{cand} 无法正向打在 environment/repo/ 上(树与参考补丁不匹配)") elif rev: fail(pkg, f"{cand} 反向也能打上 —— environment/repo/ 似乎已是修好的状态") break for dp, dn, fn in os.walk(root): for f in list(dn) + list(fn): rel = os.path.relpath(os.path.join(dp, f), root) if JUNK.search(rel): fail(pkg, f"残留构建/编辑器产物: {rel}") check_runnable(pkg, root, meta) tasks = [(b, t) for b in WANT for t in sorted(os.listdir(os.path.join(TASKS, b))) if not t.startswith("_") and not t.startswith("delete_") and os.path.isdir(os.path.join(TASKS, b, t))] for b, t in tasks: check_pkg(b, t) print(f"fai_bench 自检 — 根目录 {ROOT}") print(f" 子集 {', '.join(WANT)} / 共 {len(tasks)} 题") if warns: print(f"\n提示 {len(warns)} 条:") for w in warns[:40]: print(f" · {w}") if len(warns) > 40: print(f" · … 另 {len(warns) - 40} 条") if fails: print(f"\n失败 {len(fails)} 条:") for f in fails: print(f" ✗ {f}") sys.exit(1) print(f"\n全部通过:{len(tasks)} 题,0 失败。")