kali20gakki/msAgent / skills/accuracy/rl-consistency-dump-analysis
rl-consistency-dump-analysis
Verified通用训推一致性(prefill)的 dump 数据分析 skill。分析**已采集**的训推两侧 dump:探查 dump 结构 → 语义边界映射(不要求算子同名)→ 逐边界精确门禁比对 → 定位首差异 → ULP 量化区分舍入/逻辑级 → 追放大链 → 差异点隔离验证(打桩迭代:输出打桩指令给采集侧,等它用 msprobe `load` 覆盖边界再采后比下游,枚举所有独立差异点)→ 排除根因假设 → 输出整体结论 → 产出分析报告。与具体框架/硬件解耦,适用 msprobe PrecisionDebugger 或等价工具采的任何 dump。触发词:训推一致性分析、dump 数据分析、dump 已采好帮我分析、首差异定位、所有差异点、差异点隔离、打桩指令、对齐报告、训推不一致结论、比较两侧 tensor。采集与打桩再采由 rl-prefill-consistency-dump 负责,本 skill 只消费已落盘 dump、输出打桩指令驱动再采。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 19.5 KB |
| references/analysis.md | markdown | 4.8 KB |
| references/dump-structure.md | markdown | 3.9 KB |
| references/pitfalls.md | markdown | 3.2 KB |
| references/prefill-logprob-diff.md | markdown | 4.2 KB |
| references/report-source.md | markdown | 5.3 KB |
| references/semantic-mapping.md | markdown | 5.2 KB |
| references/stubbing-loop.md | markdown | 9.0 KB |
| references/task_card.json | config | 1.6 KB |
| references/templates/alignment_report_source.zh-CN.json | config | 3.6 KB |
| references/templates/stub_instruction.example.json | config | 1.2 KB |
| scripts/analyze_dump_boundaries.py | script | 12.6 KB |
| scripts/generate_alignment_report.py | script | 87.2 KB |
| scripts/probe_dump.py | script | 9.8 KB |
references/templates/stub_instruction.example.json
View on GitHub{
"round": 1,
"source_side": "infer",
"target_side": "train",
"isolate_boundary": "layer0.attn.q_proj_output",
"side_operators": {
"train": "Module.<训练侧模块路径>.<训练侧算子类名>.forward.<N>",
"infer": "Module.<推理侧模块路径>.<推理侧算子类名>.forward.<N>"
},
"source_dump": "<源侧 dump 根>/<step>/<rank>/dump_tensor_data",
"target_run": "<目标侧运行描述,如 train/infer 及卡号>",
"modules": [
{
"target_entry": "Module.<训练侧模块路径>.<训练侧算子类名>.forward.<N>",
"source_files": [
"Module.<推理侧模块路径>.<推理侧算子类名>.forward.<N>.output.<n>.pt"
],
"transform": "融合输出按权重行序取 q 段: x[:, :num_heads*head_dim] → squeeze 前导 batch 维 → 转 bf16; 融合布局(q|kv 还是 kv|q)现场 bitwise 验证后定切片",
"dtype": "bf16",
"note": "q_dim / kv_dim 与融合布局以现场语义映射为准, 勿沿用旧结论"
}
],
"dump_after_load": true,
"expected": "覆盖后 train 侧 q_proj 输入与 infer 侧 q 段 bitwise 一致; 若下游仍逻辑级差异 → 确认下一独立差异",
"stop_if": "下游全 EXACT → 结束枚举; 只剩余舍入级 → 记已知 gap"
}
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 |