Skip to content

Ascend/msagent / skills/profiler/msagent-profiler-breakdown

msagent-profiler-breakdown

Verified

DB格式Profiling数据场景化性能拆解:自动完成场景确认(框架/任务类型判定)、阶段拆解与单次执行边界拆解,并给出报告。Use when 需要对昇腾 Profiling DB 做阶段拆解、step 归因或生成性能报告, or when the user says "帮我按任务阶段拆解这个Profiling数据", "多层次性能耗时拆解", "分析这个 vllm/sglang/verl 数据,主要耗时都花费在哪里了"。

references/breakdown-schema.md

View on GitHub
# 拆解树节点契约(原 schemas/breakdown.schema.json)

> 本契约描述 msagent-profiler-breakdown(decompose.py)产出的分级别耗时节点(阶段 / 单次执行),
> 以及后续 layer-decompose 消费时的字段口径。层级用 `level` 字段表达(不靠目录),
> 结论阶梯深度见设计方案 §3.7。

## level 枚举

| level  | 语义                                                                                        | 由谁产出                                                                |
| ------ | ----------------------------------------------------------------------------------------- | ------------------------------------------------------------------- |
| `端到端`  | 会话整体跨度                                                                                    | msagent-profiler-breakdown(breakdown.json 顶层)                       |
| `阶段`   | 推理五 stage(prepare\_input / forward / post\_process / sample\_token / draft\_token)或 RL 阶段 | msagent-profiler-breakdown(breakdown.json children)                 |
| `单次执行` | 一次 forward(decode step)                                                                   | msagent-profiler-breakdown(breakdown.json children 中 level=单次执行 分支) |
| `单层`   | transformer 第 N 层                                                                         | layer-decompose(暂缓,经 --framework 继承场景)                              |
| `层内`   | 层内子模块(attention / MLP(MoE) / 通信 / 归一化 / 量化反量化)                                            | layer-decompose(暂缓)                                                 |

## 节点通用字段

| 字段                 | 类型           | 语义                                                       |
| ------------------ | ------------ | -------------------------------------------------------- |
| `level`            | string(上表枚举) | 拆解层级                                                     |
| `name`             | string       | 节点名:阶段名 / step 序号 / 层名 / 子模块名                            |
| `start` / `end`    | number(ns)   | 起始 / 结束时间                                                |
| `wall_time`        | number(ns)   | 墙钟耗时 = end - start                                       |
| `self_time`        | number(ns)   | 净自耗 = wall\_time - Σ children.wall\_time(主排序用)           |
| `overlap`          | number(ns)   | 并行/重叠收益 = wall\_time - self\_time(单独列示)                  |
| `children`         | array        | 子节点(递归本契约)                                               |
| `confidence`       | object       | `{method(A/B/C), level(high/medium/low), detail}`        |
| `conservation_gap` | number(ns)   | 守恒缺口 = wall\_time - Σ children.self\_time                |
| `stats`            | object       | 同类节点跨 occurrence 聚合时填充:count/p50/p99/min/max/variance/cv |
| `anomaly`          | object       | 异常/漂移标记(is\_outlier / is\_drift / threshold\_ref)        |

> 阶段节点(msagent-profiler-breakdown 的 breakdown.json children 中 level=阶段)额外带 `count`(次数)、`wall_mean` / `wall_max` / `wall_min`(单次均值/最大/最小)、`first`(首个起止/wall),供报告展示阶段统计;`--inspect-index` 时再带 `inspect`。

## confidence

- `method`:A = 框架打点 / B = 注入打点 / C = 数据推测 / `with_modules` / `sequence_align`。

- `level`:`high` / `medium` / `low`(判据见 check-rules.md 与设计方案 §3.7)。

## 约束

- msagent-profiler-breakdown 输出为独立 JSON(breakdown.json 统一拆解树,stage.json 为同内容别名;layers.json 由 layer-decompose 后续产出),**不经过统一事件 JSON 中间层**,
  数据源为昇腾 Profiling DB 直查(见 `references/db-usage.md`)。

- 字段口径以本契约为准;scripts 仅用 Python 标准库,不依赖 JSON Schema 校验库。

slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall