kali20gakki/msAgent / skills/accuracy/rl-prefill-consistency-dump
rl-prefill-consistency-dump
Verified训推一致性(prefill)采集侧:为训练/推理两侧采集可比的 prefill dump(输入对齐 PROMPTS_ONLY → 参数对齐校验(dump 前前置:校验两侧 config 超参/结构/权重,不一致输出参数不一致点并对齐)→ msprobe 接入采集 → 采集验证 → 恢复代码)。按用户需求判断用法:① 仅采集 dump——用户只要 dump 数据,走采集主流程(0→7→9),直接产出两侧可比 dump;② 配合分析——用户要求找所有独立差异点/差异点隔离验证/配合 rl-consistency-dump-analysis 分析时,才按分析侧打桩指令用 msprobe `load` 覆盖边界二次采集(阶段 8)再恢复代码。差异分析、首差异定位、打桩指令、整体结论由 rl-consistency-dump-analysis 承担。触发词:训推一致、训推一致性、prefill 比对、msprobe dump/采集、精度比对数据采集、打桩二次采集、参数对齐校验、参数不一致点、mHC 配置不一致、权重不同步、清卡验环境后采数据;已采 dump 的分析转 rl-consistency-dump-analysis。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 25.7 KB |
| references/msprobe-config.md | markdown | 5.4 KB |
| references/new-framework.md | markdown | 5.2 KB |
| references/param-alignment.md | markdown | 14.4 KB |
| references/pitfalls.md | markdown | 22.1 KB |
| references/prefill-logprob-diff.md | markdown | 5.6 KB |
| references/probs-diff-monitoring.md | markdown | 6.6 KB |
| references/slime.md | markdown | 5.4 KB |
| references/stubbing.md | markdown | 7.7 KB |
| references/task_card.json | config | 3.2 KB |
| references/templates/param_alignment_instruction.example.json | config | 1.4 KB |
| references/templates/stub_instruction.example.json | config | 1.2 KB |
| references/verl.md | markdown | 5.0 KB |
| references/vime.md | markdown | 9.5 KB |
| scripts/param_align_check.py | script | 8.7 KB |
| scripts/param_align_provider_dump.py | script | 5.7 KB |
| scripts/param_alignment_mapping.example.json | config | 2.3 KB |
scripts/param_align_check.py
View on GitHub#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""参数对齐校验(阶段 5)静态 config 比对脚本 —— 通用版, 不绑定任何具体模型。
读训练侧实际生效 config 与推理侧 hf_config, 逐项比对输出 param_alignment_report.json。
两种模式:
1. 传 --mapping <字段映射 json>: 按映射逐项语义比对(两侧字段名可不同, 支持嵌套点分路径
与结构断言 infer_expected)。映射写法见 param_alignment_mapping.example.json。
2. 不传 --mapping: 自动同名键比对(零配置兜底)——枚举两侧顶层键并集, 同名键直接比,
单侧缺失标 MISSING_SIDE, 嵌套 dict 值标 SKIP。适用快速筛查与建映射前的摸底。
纯 Python, 无 torch 依赖, 可在任意容器/宿主跑(权重/结构比对是动态层, 见 param-alignment.md §5.3)。
用法:
python3 param_align_check.py \
--train-cfg <训练侧实际生效 config.json(provider/bridge 转换结果)> \
--infer-cfg <推理侧 config.json> \
[--mapping <字段映射 json>] [--model <模型标识>] [--out <报告路径>]
注意:训练侧 config 必须是"实际转换结果", 不能只读原始 config.json——转换字段名写错时
原始文件有值但训练侧没接住, 只比原始文件会漏。
"""
import argparse
import json
import sys
def load_cfg(path):
with open(path, encoding="utf-8") as f:
return json.load(f)
def get(cfg, field):
"""取字段值; 支持嵌套点分路径。字段不存在返回 (None, True)。field 为 None 表示语义上没有该字段。"""
if field is None:
return None, False
cur = cfg
for part in field.split("."):
if isinstance(cur, dict) and part in cur:
cur = cur[part]
else:
return None, True
return cur, False
def compare(a, b):
"""比较两侧值: 标量/字符串/等长 list 逐元素。返回 (equal, detail)。
不可比(嵌套 dict 等)返回 (None, 原因), 由调用方标 SKIP。
bool 与数值/字符串类型不一致时显式判不等——True==1、1==1.0 在 Python 中为真,
但两侧 config 分属不同框架时 bool/int 表述差异应暴露为 DIFF, 不能静默算 EXACT。"""
if a is None or b is None:
return a is b, f"{a} vs {b}"
if isinstance(a, bool) or isinstance(b, bool):
return (isinstance(a, bool) and isinstance(b, bool) and a == b), f"{a!r} vs {b!r}"
if isinstance(a, (int, float)) and isinstance(b, (int, float)):
return a == b, f"{a} vs {b}"
if isinstance(a, str) and isinstance(b, str):
return a == b, f"{a} vs {b}"
if isinstance(a, list) and isinstance(b, list):
if len(a) != len(b):
return False, f"list len {len(a)} vs {len(b)}"
diffs = [(i, x, y) for i, (x, y) in enumerate(zip(a, b)) if x != y]
return not diffs, f"list len={len(a)}, first_diff={diffs[:3]}"
if isinstance(a, dict) or isinstance(b, dict):
return None, "嵌套 dict 值, 非标量不自动比(映射模式可指定点分路径深比)"
return a == b, f"{a} vs {b}"
def compare_item(train_cfg, infer_cfg, semantic, tfield, ifield, infer_expected=None):
"""单项比对。返回 item dict。"""
tv, tv_missing = get(train_cfg, tfield)
iv, iv_missing = get(infer_cfg, ifield)
if tfield is None and ifield is None:
status, detail = "N/A", "两侧均无对应字段"
elif infer_expected is not None:
# 结构断言: 一侧无独立字段(该侧恒启用/恒为某值), 用期望值校验另一侧
equal = (tv == infer_expected) and not tv_missing
status = "EXACT" if equal else "DIFF"
detail = f"train={tv} expected={infer_expected}"
elif tfield is None or ifield is None:
# 一侧在语义上没有对应字段: 标 MISSING_SIDE(不是值不一致), 查该侧代码是否硬编码同值
side = "train" if tfield is None else "infer"
status = "MISSING_SIDE"
detail = f"{side} 侧无对应字段; 查该侧代码是否硬编码同值"
elif tv_missing or iv_missing:
status = "MISSING_SIDE"
detail = f"train_missing={tv_missing} infer_missing={iv_missing}"
else:
equal, detail = compare(tv, iv)
if equal is None:
status = "SKIP"
else:
status = "EXACT" if equal else "DIFF"
return {
"semantic": semantic,
"train": {"field": tfield, "value": tv, "missing": tv_missing},
"infer": {"field": ifield, "value": iv, "missing": iv_missing},
"status": status,
"detail": detail,
}
def build_mapping_items(train_cfg, infer_cfg, mapping):
"""按映射生成比对条目(带 note)。"""
for it in mapping.get("items", []):
item = compare_item(
train_cfg, infer_cfg,
it["semantic"], it.get("train"), it.get("infer"), it.get("infer_expected"),
)
item["note"] = it.get("note", "")
yield item
def build_auto_items(train_cfg, infer_cfg):
"""无映射时自动同名键比对: 枚举两侧顶层键并集。"""
keys = list(dict.fromkeys([*train_cfg.keys(), *infer_cfg.keys()]))
items = []
skipped = []
for k in keys:
item = compare_item(
train_cfg, infer_cfg,
k, k if k in train_cfg else None, k if k in infer_cfg else None, None,
)
if item["status"] == "SKIP":
skipped.append(k)
items.append(item)
return items, skipped
def run(train_cfg, infer_cfg, mapping):
if mapping is not None:
items = list(build_mapping_items(train_cfg, infer_cfg, mapping))
prefix = "P"
# 映射模式也要收集 SKIP(嵌套 dict 值)/N/A(两侧均无字段)条目,否则这些未比项被静默吞掉,
# 汇总会误报"全部 EXACT"(假成功)。
skipped = [it["semantic"] for it in items if it["status"] in ("SKIP", "N/A")]
mode = "mapping"
else:
items, skipped = build_auto_items(train_cfg, infer_cfg)
prefix = "A"
mode = "auto"
for i, it in enumerate(items):
it["id"] = f"{prefix}{i + 1}"
mismatch = [it["id"] for it in items if it["status"] in ("DIFF", "MISSING_SIDE")]
return {"mode": mode, "items": items, "skipped": skipped, "mismatch_points": mismatch}
def main():
ap = argparse.ArgumentParser(description="参数对齐校验(阶段 5)静态 config 比对 —— 通用版")
ap.add_argument("--train-cfg", required=True, help="训练侧实际生效 config.json(provider/bridge 转换结果)")
ap.add_argument("--infer-cfg", required=True, help="推理侧 config.json")
ap.add_argument("--mapping", default=None,
help="字段映射 json(写法见 param_alignment_mapping.example.json); 不传则自动同名键比对")
ap.add_argument("--model", default="unknown", help="模型标识(写入报告, 映射文件有 model 则优先)")
ap.add_argument("--out", default=None, help="输出 param_alignment_report.json 路径(默认 stdout)")
args = ap.parse_args()
train_cfg = load_cfg(args.train_cfg)
infer_cfg = load_cfg(args.infer_cfg)
mapping = load_cfg(args.mapping) if args.mapping else None
res = run(train_cfg, infer_cfg, mapping)
model = (mapping.get("model") or args.model) if mapping is not None else args.model
mismatches = res["mismatch_points"]
skipped = res["skipped"]
if mismatches:
summary = f"参数对齐: {len(mismatches)} 个不一致点 {mismatches}, 需对齐后采"
elif skipped:
# 有 SKIP/N-A 未比项时不能宣称"全部 EXACT",仅值一致但未全覆盖,属待确认而非可采。
summary = (f"参数对齐: 无值不一致, 但 {len(skipped)} 个条目未比"
f"(嵌套 dict 值或两侧均无字段, 见 skipped_keys), 非全部 EXACT, 建议补映射确认后再采")
else:
summary = "参数对齐: 全部 EXACT, 可采"
if skipped and mismatches:
summary += f"; 另有 {len(skipped)} 个条目未比(见 skipped_keys, 可补映射深比)"
report = {
"model": model,
"mode": res["mode"],
"checked_at": "<现场填写>",
"items": res["items"],
"skipped_keys": res["skipped"],
"mismatch_points": res["mismatch_points"],
"weights": {"status": "N/A(动态层, 见 param-alignment.md §5.3)"},
"structure": {"status": "N/A(动态层, 见 param-alignment.md §5.3)"},
"summary": summary,
}
if args.out:
with open(args.out, "w", encoding="utf-8") as f:
json.dump(report, f, ensure_ascii=False, indent=2)
print(f"written: {args.out}")
else:
print(json.dumps(report, ensure_ascii=False, indent=2))
print(summary)
sys.exit(0 if not res["mismatch_points"] else 1)
if __name__ == "__main__":
main()
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |