ops-tune-workbench-plugin 0.0.0-stage → 0.0.1-beta.2

package/skills/tool-ops-tune/ops-collect/scripts/isolation_test.py

@@ -0,0 +1,454 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# ----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# ----------------------------------------------------------------------------------------------------------
+"""isolation_test.py — MC² 隔离测试编排（T_comm / T_compute 采集 → R 值）
+
+设计对齐（docs/算子调优AGENT设计文档.md §4.3 数据集模型）：
+  - MC² 隔离测试方法论此前长期停留在文档形态（复制源码、注释 process、tileCnt=1、
+    多 rank 采集、算 R 值全靠人工执行）—— 本脚本把它变成可重复的实验编排。
+  - 纪律继承：隔离测试在源码副本进行（不碰用户原目录）；多 rank 并行采集；
+    循环 ≥10 次前 5 warm-up；跨 rank 取 max（木桶效应）。
+  - 输出直接可喂 verdict.py：--scalar t_comm_us=...,t_compute_us=...
+
+用法（典型，宿主程序须支持 COMM_ONLY/COMPUTE_ONLY 环境变量开关或源码注释由 Agent 完成）：
+  python3 isolation_test.py --exe ./build/mc2_demo --args "4096 4096 4096 4 perf 1" \\
+      --prof-dir ./prof/iso --output iso_result.json
+
+输出 iso_result.json：
+  { t_comm_us, t_compute_us, r_value, bound_hint, rank_detail: {...}, steady_rule: "10轮前5warmup后5均值，跨rank取max" }
+
+R 值档位真源：ops-verdict 规则包 L3-inter-card.yaml 的 r_value 叶子条件（脚本启动时解析，
+解析失败即报错退出——fail loud，保证规则演进时脚本不会静默回退到旧常量）。
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+WARMUP = 5  # 前 5 次 warm-up（L2 预热、DVFS 爬升）
+LOOPS = 10  # 循环至少 10 次
+# MTE2 占比候选列（与 L1 词汇表 mte2_ratio 同源，三代列名兼容）
+MTE2_RATIO_COLUMNS = [
+    "aic_mte2_ratio",
+    "aiv_mte2_ratio",
+    "ai_mte2_ratio",
+    "Mte2Utilization(%)",
+    "mte2_ratio",
+    "MTE2(%)",
+]
+
+# L3 规则包默认路径（相对本脚本定位，仓内布局免传参；--rules 可覆写）
+L3_RULES_FILE = "L3-inter-card" + ".yaml"  # MC² 卡间判定规则包
+DEFAULT_L3_RULES = (
+    Path(__file__).resolve().parents[2] / "ops-verdict" / "rules" / L3_RULES_FILE
+)
+
+
+def r_bands_from_rules(rules_path: str) -> list[tuple]:
+    """从 L3 规则包 YAML 派生 R 值档位表（规则 YAML 是唯一真源）。
+
+    解析每条以 r_value 为唯一 metric 叶子条件、区间形状完整的规则，产出
+    [(lo, hi, rule_id, name), ...] 按 hi 升序（lo=None 表示无下界）。
+    解析不出任何档位 → 直接抛错退出（fail loud）。
+    """
+    import yaml
+
+    pack = yaml.safe_load(Path(rules_path).read_text(encoding="utf-8"))
+    bands = []
+    for rule in pack.get("rules", []):
+        leaves = _collect_metric_leaves(rule.get("condition"), "r_value")
+        if not leaves:
+            continue
+        lo = hi = None
+        ok = True
+        for leaf in leaves:
+            op, val = leaf["op"], leaf["value"]
+            if op == "between" and isinstance(val, list) and len(val) == 2:
+                lo, hi = float(val[0]), float(val[1])
+            elif isinstance(val, (int, float)):
+                if op in ("gt", "gte") and lo is None:
+                    lo = float(val)
+                elif op in ("lt", "lte") and hi is None:
+                    hi = float(val)
+                else:
+                    ok = False
+                    break
+            else:
+                ok = False
+                break
+        if ok and (lo is not None or hi is not None):
+            # 单叶 gt（无上界，如 STRONG_COMPUTE R>2）：hi 置 inf，作为兜底档
+            bands.append(
+                (
+                    lo,
+                    hi if hi is not None else float("inf"),
+                    rule["id"],
+                    rule.get("name", rule["id"]),
+                )
+            )
+    if not bands:
+        raise RuntimeError(
+            f"无法从 {rules_path} 解析出 r_value 档位（规则结构变化？fail loud 退出）"
+        )
+    return sorted(bands, key=lambda b: b[1])
+
+
+def classify_r(r: float, bands: list[tuple]) -> tuple[str, str]:
+    """按派生档位分类 R 值：唯一命中档 = 唯一满足区间条件的档；多重命中取最紧。
+
+    边界对齐规则表语义：STRONG_COMM 为开区间上界（R<hi），其余档下界 gte /
+    上界 lte/between——r 恰落档界（R=0.5/0.9/1.1/2.0）时同时满足两档的
+    "r ≤ hi"宽松判定，须比较真实区间条件：开上界档要求 r < hi。
+    """
+    # 第一轮：严格区间命中（lo ≤ r 且 r < hi 开上界 / r ≤ hi 闭上界）
+    strict = []
+    for lo, hi, bid, name in bands:
+        if (lo is None or r >= lo) and r < hi:
+            strict.append((lo, hi, bid, name))
+        elif (lo is None or r >= lo) and r == hi:
+            # r 恰等于闭上界档（between/lte）的 hi → 该档命中（如 BALANCED [0.9,1.1] 命中 R=1.1）
+            # 与开上界档（lt）区分：开上界在 r==hi 时不命中
+            strict.append((lo, hi, bid, name)) if _has_closed_upper(
+                bid, bands
+            ) else None
+    if len(strict) == 1:
+        return strict[0][2], strict[0][3]
+    if len(strict) > 1:
+        # 多重命中取最紧上界档（区间重叠时的确定性规则；正常规则集不触发）
+        strict.sort(key=lambda b: b[1])
+        return strict[0][2], strict[0][3]
+    # 无严格命中（r 落在档界间隙或超出覆盖）→ 兜底最紧上界
+    for lo, hi, bid, name in bands:
+        if r <= hi:
+            return bid, name
+    return "UNRESOLVED", "超出规则档位覆盖范围（检查 L3 规则包区间连续性）"
+
+
+def _has_closed_upper(bid: str, bands: list[tuple]) -> bool:
+    """该档上界是否为闭（between/lte）。当前规则集：STRONG_COMM（lt）与
+    COMM_BOUND（lt 上界）为开，BALANCED/COMPUTE_BOUND 为闭。按 id 约定
+    强通信/通信档上界开、其余闭——由 r_bands_from_rules 的区间形状保证。"""
+    return bid not in ("L3_MC2_STRONG_COMM", "L3_MC2_COMM_BOUND")
+
+
+def _collect_metric_leaves(cond, metric: str) -> list[dict]:
+    """递归收集条件树中引用指定 metric 的叶子。"""
+    out = []
+    if cond is None:
+        return out
+    if isinstance(cond, list):
+        for sub in cond:
+            out.extend(_collect_metric_leaves(sub, metric))
+        return out
+    if isinstance(cond, dict):
+        if "metric" in cond and cond["metric"] == metric:
+            out.append(cond)
+        for key in ("all", "any"):
+            if key in cond:
+                out.extend(_collect_metric_leaves(cond[key], metric))
+    return out
+
+
+def parse_task_durations(prof_dir: str) -> dict[str, list[float]]:
+    """从 msprof 产出的 op_summary_*.csv 提取 Task Duration(us)，按文件分组返回。
+
+    每个文件 ≈ 一个 rank 的独立 op_summary（msprof 每设备一份 PROF 目录）——
+    分组是「逐 rank 稳态均值再跨 rank 取 max」（木桶口径）的前提（检视 F19：
+    全部行统一求均值在数学上不等价，mean(所有行) ≤ max(各 rank 均值)，
+    系统性低估 T → R 偏向「计算 bound」结论）。
+    返回 {文件名: [时长列表]}；空文件不产生键。
+    """
+    groups: dict[str, list[float]] = {}
+    for p in sorted(Path(prof_dir).rglob("op_summary_*.csv")):
+        vals = []
+        with open(p, encoding="utf-8-sig", errors="replace") as f:
+            for row in csv.DictReader(f):
+                v = row.get("Task Duration(us)") or row.get("Task Duration")
+                if v:
+                    try:
+                        vals.append(float(str(v).strip()))
+                    except ValueError:
+                        pass
+        if vals:
+            groups[str(p)] = vals
+    return groups
+
+
+def steady_max(grouped: dict[str, list[float]], warmup: int) -> float | None:
+    """木桶口径稳态聚合（检视 F19）：每 rank（文件组）内丢弃前 warmup 次取均值，
+    再跨 rank 取 max——所有 rank 经 BarrierAll/CrossCoreFlag 同步，整体性能由
+    最慢 rank 决定。"""
+    per_rank = []
+    for vals in grouped.values():
+        steady = vals[warmup:] if len(vals) > warmup else vals
+        if steady:
+            per_rank.append(sum(steady) / len(steady))
+    return max(per_rank) if per_rank else None
+
+
+def parse_pipe_ratio(prof_dir: str, columns: list[str]) -> dict[str, float]:
+    """从 op_summary_*.csv 提取 pipe 占比列的逐 rank 稳态均值（F8：mte2_ratio_pure 等的生产者）。
+
+    返回 {文件名: 稳态均值}；列不存在或全空 → 空 dict。
+    """
+    groups: dict[str, float] = {}
+    for p in sorted(Path(prof_dir).rglob("op_summary_*.csv")):
+        vals = []
+        with open(p, encoding="utf-8-sig", errors="replace") as f:
+            for row in csv.DictReader(f):
+                for col in columns:
+                    v = row.get(col)
+                    if v and str(v).strip() not in ("", "N/A", "NA", "-"):
+                        try:
+                            fv = float(str(v).strip().replace("%", ""))
+                            vals.append(
+                                fv * 100 if 0 <= fv <= 1.5 else fv
+                            )  # ratio_to_pct 兼容
+                            break
+                        except ValueError:
+                            pass
+        if vals:
+            groups[str(p)] = sum(vals) / len(vals)
+    return groups
+
+
+def parse_switch(spec: str) -> dict:
+    """把 --comm-switch/--compute-switch 解析成环境变量注入 dict。
+
+    支持两种形式：
+      - 纯 KEY=VALUE 或 KEY=VALUE,K2=V2：注入这些变量（默认 MC2_MODE=COMM_ONLY 等）
+      - 裸字符串（无 =）：视为提示文本，不注入（打印用）
+    """
+    env = {}
+    for kv in spec.split(","):
+        kv = kv.strip()
+        if "=" in kv:
+            k, v = kv.split("=", 1)
+            env[k.strip()] = v.strip()
+    return env
+
+
+def run_msprof_task_based(
+    exe: str, exe_args: str, out: str, env_extra: dict
+) -> dict[str, list[float]]:
+    """msprof task-based 采集一轮（MC² 专用路径），返回按 op_summary 文件（≈按 rank）
+    分组的 Task Duration。
+
+    env 采用整体替换语义（subprocess.run(env=...) 不继承父进程环境），因此必须
+    以 os.environ 为底、env_extra 覆盖——否则 PATH/ASCEND_HOME_PATH/LD_LIBRARY_PATH
+    全部丢失，msprof 无法运行（检视 F2）。
+    """
+    env = {**os.environ, **env_extra}
+    cmd = [
+        "msprof",
+        f"--output={out}",
+        "--ai-core=on",
+        "--aic-mode=task-based",
+        "--aic-metrics=PipeUtilization",
+        "--task-time=on",
+        "--ascendcl=on",
+        f"--application={exe} {exe_args}",
+    ]
+    proc = subprocess.run(cmd, capture_output=True, text=True, timeout=1800, env=env)
+    if proc.returncode != 0:
+        raise RuntimeError(f"msprof 失败: {proc.stderr[:400]}")
+    return parse_task_durations(out)
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description="MC² 隔离测试编排（T_comm/T_compute → R）")
+    ap.add_argument(
+        "--exe", required=True, help="待测可执行文件（须按 rank 数自行 fork）"
+    )
+    ap.add_argument(
+        "--args", default="", help='程序参数（如 "M K N rankNum perf tileCnt"）'
+    )
+    ap.add_argument(
+        "--comm-switch",
+        default="MC2_MODE=COMM_ONLY",
+        help="纯通信开关（K=V 或 K1=V1,K2=V2 环境变量形式，注入采集子进程；"
+        "源码注释模式由 Agent 在副本完成后调用本脚本 --from-prof 收敛）",
+    )
+    ap.add_argument(
+        "--compute-switch",
+        default="MC2_MODE=COMPUTE_ONLY",
+        help="纯计算开关（同上形式，注入采集子进程）",
+    )
+    ap.add_argument(
+        "--switch-mode",
+        default="env",
+        choices=["env", "from-prof"],
+        help="env=程序自带模式开关；from-prof=已手工注释的两份 profiling 目录直接给路径",
+    )
+    ap.add_argument(
+        "--comm-prof", default=None, help="from-prof 模式：纯通信 profiling 目录"
+    )
+    ap.add_argument(
+        "--compute-prof", default=None, help="from-prof 模式：纯计算 profiling 目录"
+    )
+    ap.add_argument("--prof-dir", default="./prof/iso", help="env 模式采集落盘根目录")
+    ap.add_argument(
+        "--rules",
+        default=str(DEFAULT_L3_RULES),
+        help="L3 规则包 YAML（R 值档位唯一真源；默认仓内 L3-inter-card.yaml）",
+    )
+    ap.add_argument("--loops", type=int, default=LOOPS)
+    ap.add_argument("--warmup", type=int, default=WARMUP)
+    ap.add_argument("--output", required=True)
+    args = ap.parse_args()
+
+    if args.switch_mode == "from-prof":
+        comm_groups = parse_task_durations(args.comm_prof)
+        compute_groups = parse_task_durations(args.compute_prof)
+        if not comm_groups or not compute_groups:
+            print(
+                "ERROR: from-prof 目录中未找到 op_summary Task Duration 数据",
+                file=sys.stderr,
+            )
+            return 1
+        t_comm, t_compute = (
+            steady_max(comm_groups, args.warmup),
+            steady_max(compute_groups, args.warmup),
+        )
+        t_full, full_groups, mte2_ratio_full = None, {}, None
+        detail = {
+            "mode": "from-prof",
+            "comm_prof": args.comm_prof,
+            "compute_prof": args.compute_prof,
+        }
+        # F8：纯计算路径 pipe 占比（mte2_ratio_pure 的生产者——隔离计算 run 的
+        # aic_mte2_ratio 稳态均值，跨 rank 取 max，与 T 口径一致）
+        compute_pure_ratio = parse_pipe_ratio(args.compute_prof, MTE2_RATIO_COLUMNS)
+        mte2_ratio_pure = (
+            round(max(compute_pure_ratio.values()), 1) if compute_pure_ratio else None
+        )
+    else:
+        # env 开关模式：--comm-switch/--compute-switch 支持 K=V（或 K1=V1,K2=V2）形式，
+        # 解析出的变量真正注入子进程 env（消灭死参数，检视 F2/N2）
+        comm_env = parse_switch(args.comm_switch) or {"MC2_MODE": "COMM_ONLY"}
+        compute_env = parse_switch(args.compute_switch) or {"MC2_MODE": "COMPUTE_ONLY"}
+        comm_out = str(Path(args.prof_dir) / "comm")
+        compute_out = str(Path(args.prof_dir) / "compute")
+        print(f"[iso] 采集纯通信（{args.comm_switch}）…")
+        comm_groups = run_msprof_task_based(args.exe, args.args, comm_out, comm_env)
+        print(f"[iso] 采集纯计算（{args.compute_switch}）…")
+        compute_groups = run_msprof_task_based(
+            args.exe, args.args, compute_out, compute_env
+        )
+        print("[iso] 采集融合态（全量，F8：mte2_ratio_full / mask_ratio 生产者）…")
+        full_out = str(Path(args.prof_dir) / "full")
+        full_groups = run_msprof_task_based(args.exe, args.args, full_out, {})
+        t_comm, t_compute = (
+            steady_max(comm_groups, args.warmup),
+            steady_max(compute_groups, args.warmup),
+        )
+        t_full = steady_max(full_groups, args.warmup)
+        detail = {
+            "mode": "env",
+            "comm_ranks": len(comm_groups),
+            "compute_ranks": len(compute_groups),
+            "full_ranks": len(full_groups),
+        }
+        compute_pure_ratio = parse_pipe_ratio(compute_out, MTE2_RATIO_COLUMNS)
+        mte2_ratio_pure = (
+            round(max(compute_pure_ratio.values()), 1) if compute_pure_ratio else None
+        )
+        full_ratio = parse_pipe_ratio(full_out, MTE2_RATIO_COLUMNS)
+        mte2_ratio_full = round(max(full_ratio.values()), 1) if full_ratio else None
+
+    if not t_comm or not t_compute:
+        print("ERROR: 稳态数据为空（检查 op_summary 行数 ≥ loops）", file=sys.stderr)
+        return 1
+
+    # F8：rank_imbalance = max(rank 稳态均值)/mean(rank 稳态均值)（融合态，跨 rank 木桶离散度）
+    rank_imbalance = None
+    if t_full is not None and full_groups:
+        rank_means = []
+        for vals in full_groups.values():
+            steady = vals[args.warmup :] if len(vals) > args.warmup else vals
+            if steady:
+                rank_means.append(sum(steady) / len(steady))
+        if len(rank_means) >= 2 and sum(rank_means) > 0:
+            rank_imbalance = round(
+                max(rank_means) / (sum(rank_means) / len(rank_means)), 3
+            )
+
+    # F8：mask_ratio = (融合前耗时 − 融合后耗时) / min(T_comm, T_compute)，上限 1.0
+    # 融合前耗时 = T_comm + T_compute（隔离态串行和）；融合后 = t_full
+    mask_ratio = None
+    if t_full is not None and t_comm and t_compute:
+        denom = min(t_comm, t_compute)
+        if denom > 0:
+            mask_ratio = round(min((t_comm + t_compute - t_full) / denom, 1.0), 3)
+
+    r = t_compute / t_comm
+    # 档位从 L3 规则包 YAML 派生（规则是唯一真源；解析失败在 r_bands_from_rules 里 fail loud）
+    bound, bound_name = classify_r(r, r_bands_from_rules(args.rules))
+
+    # F8：产出直接拼成 verdict.py 的 --scalar 串（MC² 旗舰判定链端到端可执行）
+    scalar_parts = [f"t_comm_us={t_comm:.0f}", f"t_compute_us={t_compute:.0f}"]
+    if mte2_ratio_full is not None:
+        scalar_parts.append(f"mte2_ratio_full={mte2_ratio_full:.0f}")
+    if mte2_ratio_pure is not None:
+        scalar_parts.append(f"mte2_ratio_pure={mte2_ratio_pure:.0f}")
+    if rank_imbalance is not None:
+        scalar_parts.append(f"rank_imbalance={rank_imbalance}")
+    if mask_ratio is not None:
+        scalar_parts.append(f"mask_ratio={mask_ratio}")
+    scalar_str = ",".join(scalar_parts)
+
+    result = {
+        "schema": "opstune.isolation/1.0",
+        "t_comm_us": round(t_comm, 2),
+        "t_compute_us": round(t_compute, 2),
+        "t_full_us": round(t_full, 2) if t_full is not None else None,
+        "r_value": round(r, 3),
+        "bound_hint": bound,
+        "bound_name": bound_name,
+        "mte2_ratio_pure": mte2_ratio_pure,
+        "mte2_ratio_full": mte2_ratio_full,
+        "rank_imbalance": rank_imbalance,
+        "mask_ratio": mask_ratio,
+        "verdict_scalar": scalar_str,
+        "r_bands_source": str(Path(args.rules).name),
+        "steady_rule": (
+            f"{args.loops} 轮前 {args.warmup} warm-up 后逐 rank 均值，"
+            "跨 rank 取 max（木桶效应）"
+        ),
+        "detail": detail,
+        "next": (
+            f"喂判定引擎：python3 ../ops-verdict/scripts/verdict.py "
+            f"--rules ../ops-verdict/rules/L3-inter-card.yaml,../ops-verdict/rules/L1-single-core.yaml "
+            f"--soc a5 --scalar {scalar_str} -o verdict_mc2.json"
+        ),
+    }
+    out = Path(args.output)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(
+        json.dumps(result, indent=2, ensure_ascii=False) + "\n", encoding="utf-8"
+    )
+    print(f"T_comm={t_comm:.1f}us  T_compute={t_compute:.1f}us  R={r:.3f} → {bound}")
+    if mte2_ratio_pure is not None or mte2_ratio_full is not None:
+        print(f"mte2_ratio: pure={mte2_ratio_pure} full={mte2_ratio_full}")
+    if rank_imbalance is not None:
+        print(f"rank_imbalance={rank_imbalance}  mask_ratio={mask_ratio}")
+    print(f"→ {out}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/skills/tool-ops-tune/ops-collect/scripts/jsonl_to_cases.py

@@ -0,0 +1,163 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# ----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# CANN Open Software License Agreement Version 2.0 (the "License").
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# See LICENSE in the root of the software repository for the full text of the License.
+# ----------------------------------------------------------------------------------------------------------
+"""jsonl_to_cases.py — JSONL 用例 → cases.csv 适配（REVIEW2 G11 / 反馈 P1-5①）
+
+设计对齐：
+  - cases.csv 是本仓 A1 全量铁律的输入契约（io-contracts.md §1：case_id/shape/
+    dtype/attrs），但生态内存量算子工程用 JSONL 用例（torch-ops-profiler
+    规范：每行 {"inputs": [{name/type/dtype/shape|value}...]})——两格式不通，
+    存量工程进不了 Phase 0（REVIEW2 G11 定性）。
+  - 本脚本做单向适配：JSONL → cases.csv。转换规则全落脚本（判定/转换逻辑
+    不进 LLM），歧义处 fail loud 列入 skipped 而非猜测。
+
+转换规则：
+  case_id   ：行号（c01 起）或 JSONL 行自带 case_id/name 字段（优先）
+  shape     ：首个 type=tensor 输入的 shape（多维 x 分隔）；无 tensor 输入
+              → 该行 skipped（无性能可比形态）
+  dtype     ：首个 tensor 输入的 dtype（float16→fp16 等短名归一）
+  attrs     ：type != tensor 的输入（attr 类）拼 K=V（value 为列表/布尔的
+              原样 JSON 序列化），保留原 name
+
+用法：
+  python3 jsonl_to_cases.py --jsonl layer_norm_perf_cases.jsonl -o cases.csv
+  python3 jsonl_to_cases.py --jsonl x.jsonl -o cases.csv --max-cases 20   # A1 截断同款
+"""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import sys
+from pathlib import Path
+
+DTYPE_MAP = {
+    "float16": "fp16",
+    "float32": "fp32",
+    "bfloat16": "bf16",
+    "int8": "int8",
+    "int32": "int32",
+    "int64": "int64",
+    "float": "fp32",
+    "half": "fp16",
+}
+
+
+def fmt_value(v) -> str:
+    if isinstance(v, bool):
+        return "true" if v else "false"
+    if isinstance(v, (list, tuple)):
+        return "x".join(str(x) for x in v)
+    return str(v)
+
+
+def convert_row(idx: int, row: dict) -> tuple[dict | None, str | None]:
+    """一行 JSONL → 一行 cases.csv（或 skip 原因）。"""
+    case_id = row.get("case_id") or row.get("name") or f"c{idx:02d}"
+    inputs = row.get("inputs")
+    if not isinstance(inputs, list) or not inputs:
+        return None, f"{case_id}: 无 inputs 数组"
+    tensors = [i for i in inputs if isinstance(i, dict) and i.get("type") == "tensor"]
+    if not tensors:
+        return None, f"{case_id}: 无 tensor 输入（无可比性能形态）"
+    t = tensors[0]
+    shape = t.get("shape")
+    if not isinstance(shape, (list, tuple)) or not shape:
+        return None, f"{case_id}: tensor 输入无 shape"
+    dtype = DTYPE_MAP.get(str(t.get("dtype", "")).lower())
+    if not dtype:
+        return None, f"{case_id}: 未知 dtype {t.get('dtype')!r}"
+    attrs = []
+    for i in inputs:
+        if isinstance(i, dict) and i.get("type") != "tensor" and "name" in i:
+            if "value" in i:
+                attrs.append(f"{i['name']}={fmt_value(i['value'])}")
+    return {
+        "case_id": str(case_id),
+        "shape": "x".join(str(d) for d in shape),
+        "dtype": dtype,
+        "attrs": ",".join(attrs),
+    }, None
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(
+        description="JSONL 用例 → cases.csv（A1 铁律输入适配）"
+    )
+    ap.add_argument(
+        "--jsonl", required=True, help="JSONL 用例文件（torch-ops-profiler 规范）"
+    )
+    ap.add_argument("-o", "--output", required=True, help="cases.csv 输出")
+    ap.add_argument(
+        "--max-cases",
+        type=int,
+        default=0,
+        help="截断上限（0=不截断；>20 取前 20 需告知用户，与 A1 铁律同款）",
+    )
+    args = ap.parse_args()
+
+    rows_out, skipped = [], []
+    with open(args.jsonl, encoding="utf-8") as f:
+        for idx, line in enumerate(f, 1):
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                row = json.loads(line)
+            except json.JSONDecodeError as e:
+                skipped.append(f"line {idx}: JSON 解析失败（{e}）")
+                continue
+            converted, reason = convert_row(idx, row)
+            if converted is None:
+                skipped.append(reason)
+            else:
+                rows_out.append(converted)
+
+    if not rows_out:
+        print("ERROR: 无可转换行（全部 skipped，见 stderr）", file=sys.stderr)
+        for s in skipped:
+            print(f"  ⚠ {s}", file=sys.stderr)
+        return 1
+
+    truncated = False
+    if args.max_cases and len(rows_out) > args.max_cases:
+        rows_out = rows_out[: args.max_cases]
+        truncated = True
+        print(
+            f"WARN: 用例 {len(rows_out)} > {args.max_cases}，取前 {args.max_cases}"
+            f"（A1 铁律截断，报告必须告知用户）",
+            file=sys.stderr,
+        )
+
+    out = Path(args.output)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    # csv.writer 统一转义（检视 #16 修复）：case_id/shape/dtype 来自外部 JSONL
+    # 属不可信输入，手写拼接在含逗号/引号时产生错位或非法 CSV
+    with open(out, "w", encoding="utf-8", newline="") as f:
+        writer = csv.writer(f)
+        writer.writerow(["case_id", "shape", "dtype", "attrs"])
+        for r in rows_out:
+            writer.writerow([r["case_id"], r["shape"], r["dtype"], r["attrs"]])
+
+    print(
+        f"cases.csv → {out}（{len(rows_out)} 行"
+        f"{'（A1 截断）' if truncated else ''}，skipped {len(skipped)}）"
+    )
+    for s in skipped[:10]:
+        print(f"  ⚠ skipped: {s}", file=sys.stderr)
+    if len(skipped) > 10:
+        print(f"  ⚠ …共 {len(skipped)} 条 skipped", file=sys.stderr)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())