Ascend/msagent / skills/profiler/msagent-profiler-breakdown
msagent-profiler-breakdown
VerifiedDB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"。
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 8.0 KB |
| references/breakdown-schema.md | markdown | 4.1 KB |
| references/check-rules.md | markdown | 1.5 KB |
| references/db-usage.md | markdown | 5.8 KB |
| references/frameworks/README.md | markdown | 5.5 KB |
| references/frameworks/verl.md | markdown | 11.9 KB |
| references/frameworks/vllm-instrumentation.md | markdown | 13.6 KB |
| references/frameworks/vllm-profiler-data.md | markdown | 4.6 KB |
| references/frameworks/vllm.md | markdown | 5.6 KB |
| references/report-template.md | markdown | 5.0 KB |
| references/scenario-taxonomy.md | markdown | 5.7 KB |
| resources/scenarios/inference/sglang.json | config | 549 B |
| resources/scenarios/inference/vllm.json | config | 912 B |
| resources/scenarios/rl/slime.json | config | 468 B |
| resources/scenarios/rl/verl.json | config | 1.4 KB |
| resources/scenarios/training/fsdp.json | config | 303 B |
| resources/scenarios/training/megatron.json | config | 556 B |
| resources/scenarios/training/mindspeed.json | config | 318 B |
| scripts/compare_runs.py | script | 12.4 KB |
| scripts/db_query.py | script | 12.7 KB |
| scripts/decompose.py | script | 47.7 KB |
references/frameworks/vllm-profiler-data.md
View on GitHub# vLLM / vLLM-Ascend PyTorch Profiler 采集数据与打点(torch.profiler · torch\_npu.profiler · MSTX)
> 消费 skill:`msagent-profiler-breakdown`(decompose.py,阶段归因 + 单次执行边界;prefill/decode 判别);DB 数据接入见 `references/db-usage.md`
> 性质:采集侧**事实清单**(非规则),与 `vllm-instrumentation.md`(`record_function_or_nullcontext` scope 清单)互补。
> 来源:代码静态梳理(vLLM 主仓库 + vLLM-Ascend)。**不含** MS Service Profiler、**不含**可观测性/OpenTelemetry traces 打点。
> 范围:可用于拆解的采集字段(`record_function` scope · `PYTORCH_API` / `STEP_TIME` / `CANN_API` 等 DB 字段);**不含** profiler 配置(activities / schedule / experimental\_config 等)。
> 对应原分析条目:**② PyTorch Profiler / torch\_npu.profiler 采集数据**。
***
## 1. 已核验不可用的打点通路(昇腾侧无实际采点,不再展开)
| 通路 | 昇腾实际状态 |
| --------------------- | -------------------------------------------------------------------------------------------------------------- |
| `annotate_profile`(execute\_\* 注解) | **无**:`worker.py:740` 的 `execute_model` 不打注解,且 `TorchNPUProfilerWrapper` 未覆写 `annotate_context_manager` → 返回 `nullcontext()` |
| MSTX 打点 | **默认无采点**:vLLM-Ascend `torch_npu_profiler.py:53` `msprof_tx=False`,`PYTORCH_API` 表不会出现 `type='mstx'` 记录 |
- 结论:prefill/decode 判别与 step 边界**不依赖以上通路**,改走 §2 的 `STEP_TIME` / `inputShapes` / attention kernel 类型 / `record_function` scope(见 `vllm-instrumentation.md`)。
- 若确需启用 MSTX 作精确锚点:把 `msprof_tx` 改为 `True` 重采,`PYTORCH_API.type='mstx'` 即可用。
***
## 2. 采集数据与字段 → 拆解维度映射(②「有哪些数据/字段」)
落地表结构全量见各 skill 的 `references/db-usage.md`(本 skill 关注的表/字段最小子集),本节只列**与拆解直接相关**的字段(避免重复全表)。
| 拆解维度 | 可用字段来源 | 备注 |
| ----------------- | --------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------- |
| step 边界 | `STEP_TIME.startNs/endNs`(A 策略直接来源);`record_function_or_nullcontext` scope(见 `vllm-instrumentation.md`) | 主用 `STEP_TIME` / `forward` scope |
| prefill/decode 判别 | `PYTORCH_API.inputShapes`(batch/token 维);attention kernel 类型 | Ascend 无 prefill/decode 注解,靠 shape + kernel 类型 |
| 单层归属 | `PYTORCH_API.name` + `sequenceNumber` + `fwdThreadId` + `inputShapes` 序列对齐 | module\_path 仅开模块层级采集才有 |
| 层内归因 | `PYTORCH_API.name / inputShapes / inputDtypes`(`STRING_IDS` 反查)→ 命名/数据流推断 | 见 `layer-decompose/references/layer-mapping.md` |
| 通信归因 | `CommAnalyzerTime`(wait/sync/transit/idle)/ `CommAnalyzerMatrix`(rank 对) | `analysis.db` |
| overlap 守恒 | `StepTraceTime.overlapped / communication_not_overlapped / bubble / free` | `analysis.db`,官方预拆 |
| 多卡映射 | `RANK_DEVICE_MAP`(rank→device) | PP 分层 / TP·EP 通信点 |
***
## 3. 关键澄清
1. `annotate_profile` 不是 `record_function_or_nullcontext` 通路——前者是 `torch.profiler.record_function`,后者是 `torch.autograd.profiler.record_function / nvtx / nullcontext`(详见 `vllm-instrumentation.md` §0)。
2. **Ascend 无** **`execute_*`** **注解**(`annotate_context_manager` 未覆写 → `nullcontext`),不要拿 `execute_context_*` 去昇腾 trace 里匹配;prefill/decode 改走 §2 的 shape / kernel 类型 / `STEP_TIME`。
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| 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: 197k | Updated | Claude CodeCodex | Verified | Install |
| 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: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| 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: 194k | Updated | Claude CodeCodex | Verified | Install |