Skip to content

skill-map.md

Latest commit

 

History

History
345 lines (263 loc) · 18.5 KB

skill-map.md

File metadata and controls

345 lines (263 loc) · 18.5 KB
Raw

Skill 架构地图

此详细视图由 skills/catalog.json 生成。Presentation metadata 用于人类导航；routing 字段仍是 skill ownership 合约。

数量

  • 公开：21
  • 内部 / 仅显式调用：1
  • Canonical skills 总数：22

角色与领域矩阵

Canonical ID 显示名称 角色 领域 暴露方式
model-pr-optimization-history Model and PR History planning-and-knowledge knowledge public-primary
ssh-remote-exec Remote SSH Execution gates-and-support remote-execution internal-explicit-only
tilelang-pto-kernel-workflow TileLang to PTO Kernel Workflow orchestration development public-primary
xllm-experiment-lifecycle Experiment Lifecycle orchestration lifecycle public-primary
xllm-npu-accuracy-debug Accuracy Regression Debugging analysis-and-diagnosis accuracy public-primary
xllm-npu-accuracy-runner Accuracy Evaluation Runner execution accuracy public-primary
xllm-npu-batch-perf Batch Performance Campaign orchestration performance public-primary
xllm-npu-benchmark Benchmark Fairness and Claims analysis-and-diagnosis performance public-primary
xllm-npu-build-gate Deterministic Build Gate gates-and-support build public-primary
xllm-npu-capacity-planner Serving Capacity Planner planning-and-knowledge capacity public-primary
xllm-npu-code-review NPU Code Review development-and-review development public-primary
xllm-npu-compute-simulation NPU Compute Simulation planning-and-knowledge capacity public-primary
xllm-npu-eval-runner Mixed Evaluation Orchestrator orchestration evaluation public-primary
xllm-npu-incident-triage NPU Incident Triage analysis-and-diagnosis reliability public-primary
xllm-npu-perf-runner Performance Evaluation Runner execution performance public-primary
xllm-npu-pipeline-analysis Serving Pipeline Analysis analysis-and-diagnosis profiling public-primary
xllm-npu-profiler Ascend NPU Profiler analysis-and-diagnosis profiling public-primary
xllm-npu-report-writer Evaluation Report Writer gates-and-support reporting public-primary
xllm-npu-server-manager xLLM Service Manager execution evaluation public-primary
xllm-npu-sota-loop Performance Optimization Loop orchestration performance public-primary
xllm-npu-triton-migration Triton-Ascend Operator Migration development-and-review development public-primary
xllm-npu-xllm-ops-integration xllm_ops Runtime Integration development-and-review development public-primary

依赖关系图

依赖是 catalog 声明的 canonical skill 委托边。

flowchart TD
  s00["model-pr-optimization-history"]
  s01["ssh-remote-exec"]
  s02["tilelang-pto-kernel-workflow"]
  s03["xllm-experiment-lifecycle"]
  s04["xllm-npu-accuracy-debug"]
  s05["xllm-npu-accuracy-runner"]
  s06["xllm-npu-batch-perf"]
  s07["xllm-npu-benchmark"]
  s08["xllm-npu-build-gate"]
  s09["xllm-npu-capacity-planner"]
  s10["xllm-npu-code-review"]
  s11["xllm-npu-compute-simulation"]
  s12["xllm-npu-eval-runner"]
  s13["xllm-npu-incident-triage"]
  s14["xllm-npu-perf-runner"]
  s15["xllm-npu-pipeline-analysis"]
  s16["xllm-npu-profiler"]
  s17["xllm-npu-report-writer"]
  s18["xllm-npu-server-manager"]
  s19["xllm-npu-sota-loop"]
  s20["xllm-npu-triton-migration"]
  s21["xllm-npu-xllm-ops-integration"]
  s02 --> s03
  s02 --> s07
  s02 --> s10
  s02 --> s12
  s02 --> s16
  s02 --> s21
  s03 --> s07
  s03 --> s08
  s03 --> s18
  s04 --> s05
  s05 --> s18
  s06 --> s01
  s06 --> s14
  s06 --> s17
  s06 --> s18
  s07 --> s17
  s12 --> s01
  s12 --> s05
  s12 --> s08
  s12 --> s14
  s12 --> s17
  s12 --> s18
  s13 --> s08
  s13 --> s16
  s14 --> s18
  s16 --> s15
  s18 --> s01
  s19 --> s02
  s19 --> s04
  s19 --> s07
  s19 --> s09
  s19 --> s10
  s19 --> s11
  s19 --> s12
  s19 --> s13
  s19 --> s15
  s19 --> s16
  s19 --> s20
  s19 --> s21
  s20 --> s21
  s21 --> s08
  s21 --> s12
Loading

Framework 与 backend scope

Canonical ID Framework scope Backend scope
model-pr-optimization-history xllm; vllm-ascend-dossiers; sglang-dossiers reference-artifacts
ssh-remote-exec framework-neutral remote-host; ascend-npu
tilelang-pto-kernel-workflow xllm; tilelang-generated-pto ascend-npu; tilelang; pto
xllm-experiment-lifecycle xllm; adapter-artifact-lifecycle backend-neutral-control-plane
xllm-npu-accuracy-debug xllm ascend-npu; reference-gpu
xllm-npu-accuracy-runner xllm; openai-compatible-endpoint ascend-npu
xllm-npu-batch-perf xllm ascend-npu
xllm-npu-benchmark xllm; vllm-ascend-experimental; sglang-experimental ascend-npu; nvidia-gpu-snapshot
xllm-npu-build-gate xllm; vllm-ascend-build-adapter; sglang-build-adapter ascend-npu; offline-build
xllm-npu-capacity-planner xllm; vllm-ascend-experimental; sglang-experimental ascend-npu
xllm-npu-code-review xllm ascend-npu
xllm-npu-compute-simulation xllm; vllm-ascend-experimental; sglang-experimental ascend-npu
xllm-npu-eval-runner xllm; openai-compatible-endpoint ascend-npu
xllm-npu-incident-triage xllm ascend-npu
xllm-npu-perf-runner xllm; openai-compatible-endpoint ascend-npu
xllm-npu-pipeline-analysis xllm; vllm-ascend-artifacts; sglang-artifacts ascend-npu
xllm-npu-profiler xllm; vllm-ascend-artifacts ascend-npu
xllm-npu-report-writer artifact-format-neutral backend-neutral
xllm-npu-server-manager xllm ascend-npu
xllm-npu-sota-loop xllm ascend-npu
xllm-npu-triton-migration xllm; torch-npu-ops triton-ascend; ascend-npu
xllm-npu-xllm-ops-integration xllm ascend-npu; xllm-ops

公开 Skill 详情

模型与 PR 历史（Model and PR History，model-pr-optimization-history）

查询模型 dossier 中的历史改动、风险、失败经验和后续检查。

  • 适用场景： query model optimization and PR history; retrieve prior risks validation and next checks by model keyword or code path。
  • 不适用场景： open-ended performance optimization; new benchmark execution; live incident diagnosis。
  • 输出： matching dossier sections; historical risks; recommended next checks。
  • Skill： Model and PR History

TileLang 到 PTO Kernel 优化闭环（TileLang to PTO Kernel Workflow，tilelang-pto-kernel-workflow）

执行带门禁的多轮 TileLang 与 generated PTO kernel 优化闭环。

  • 适用场景： multi-round TileLang or generated PTO kernel development and optimization; TileLang-to-PTO source and ISA specialization with production validation。
  • 不适用场景： general xllm optimization without TileLang or PTO scope; one benchmark or existing profile analysis; experiment lifecycle management; runtime integration of an already completed operator only。
  • 输出： operator baseline and profile evidence; per-round plans and decisions; validated kernel and production evidence。
  • Skill： TileLang to PTO Kernel Workflow

实验生命周期（Experiment Lifecycle，xllm-experiment-lifecycle）

创建、恢复、验证、收口并归档可审计的实验 run。

  • 适用场景： experiment run creation and recovery; checkpoint attempt evidence finalize and archive lifecycle。
  • 不适用场景： performance optimization hypothesis selection; benchmark workload execution; service or profiler operation。
  • 输出： run root; checkpoint and attempt ledger; final evidence and retention record。
  • Skill： Experiment Lifecycle

精度回归调试（Accuracy Regression Debugging，xllm-npu-accuracy-debug）

定位错误输出、评测掉分和 GPU/NPU 不一致。

  • 适用场景： accuracy regression diagnosis; wrong or garbled output diagnosis。
  • 不适用场景： accuracy measurement only; generic runtime incident。
  • 输出： minimal reproducer; root-cause evidence; validated next checks。
  • Skill： Accuracy Regression Debugging

精度评测执行器（Accuracy Evaluation Runner，xllm-npu-accuracy-runner）

对 ready 服务执行一次 EvalScope 精度 workload。

  • 适用场景： explicit EvalScope accuracy execution。
  • 不适用场景： accuracy root-cause diagnosis; mixed evaluation; benchmark fairness。
  • 输出： accuracy score; prediction records; evaluation artifacts。
  • Skill： Accuracy Evaluation Runner

批量性能 Campaign（Batch Performance Campaign，xllm-npu-batch-perf）

编排多模型、多配置或重复轮次的性能 campaign。

  • 适用场景： multi-model or multi-configuration performance campaign; repeated performance campaign。
  • 不适用场景： one ordinary performance run; fairness conclusion; accuracy evaluation。
  • 输出： campaign matrix results; per-run artifacts; aggregate summary。
  • Skill： Batch Performance Campaign

Benchmark 公平性与结论（Benchmark Fairness and Claims，xllm-npu-benchmark）

审查公平对比、SLA 候选和可发布性能结论。

  • 适用场景： fair before-after or cross-framework comparison; SLA and QPS candidate review; publishable performance conclusion。
  • 不适用场景： raw EvalScope execution only; profiling root-cause analysis。
  • 输出： fairness verdict; before/after comparison; publishable claim scope。
  • Skill： Benchmark Fairness and Claims

确定性构建门禁（Deterministic Build Gate，xllm-npu-build-gate）

验证 checkout 身份、确定性构建和 binary provenance。

  • 适用场景： explicit deterministic build and binary provenance validation。
  • 不适用场景： failed build root-cause diagnosis; performance optimization。
  • 输出： build verdict; binary provenance; dependency identity。
  • Skill： Deterministic Build Gate

Serving 容量规划（Serving Capacity Planner，xllm-npu-capacity-planner）

规划 HBM、KV cache 容量、并发和 OOM 风险。

  • 适用场景： HBM and KV-cache capacity planning; concurrency and OOM risk what-if。
  • 不适用场景： active OOM incident diagnosis; FLOPs modeling。
  • 输出： memory budget; concurrency estimate; OOM risk assessment。
  • Skill： Serving Capacity Planner

NPU 代码审查（NPU Code Review，xllm-npu-code-review）

审查 xLLM 昇腾补丁的正确性、性能风险和缺失验证。

  • 适用场景： xllm NPU-specific code review。
  • 不适用场景： generic non-NPU review; operator implementation。
  • 输出： severity-ranked findings; risk assessment; validation gaps。
  • Skill： NPU Code Review

NPU 计算量模拟（NPU Compute Simulation，xllm-npu-compute-simulation）

估算 FLOPs、MFU、硬件上限和 serving shape 成本。

  • 适用场景： FLOPs and MFU estimation; hardware lower-bound and shape what-if。
  • 不适用场景： measured benchmark conclusion; HBM capacity planning。
  • 输出： compute estimate; MFU bound; shape what-if analysis。
  • Skill： NPU Compute Simulation

混合评测编排器（Mixed Evaluation Orchestrator，xllm-npu-eval-runner）

编排服务、性能、精度和完整评测 artifacts。

  • 适用场景： mixed performance and accuracy evaluation; end-to-end evaluation artifact collection。
  • 不适用场景： service-only maintenance; benchmark fairness conclusion; profiling analysis; incident diagnosis。
  • 输出： service evidence; performance artifacts; accuracy artifacts。
  • Skill： Mixed Evaluation Orchestrator

NPU 事故诊断（NPU Incident Triage，xllm-npu-incident-triage）

复现并诊断 runtime、通信、图模式和构建事故。

  • 适用场景： runtime crash hang HCCL graph or build incident diagnosis。
  • 不适用场景： stable accuracy regression; capacity planning; healthy build gate。
  • 输出： incident classification; root-cause evidence; fix validation plan。
  • Skill： NPU Incident Triage

性能评测执行器（Performance Evaluation Runner，xllm-npu-perf-runner）

对 ready 服务执行一次 EvalScope 性能 workload。

  • 适用场景： explicit EvalScope performance execution。
  • 不适用场景： performance optimization; fair comparison; profiling analysis。
  • 输出： raw performance results; normalized metrics; run artifacts。
  • Skill： Performance Evaluation Runner

Serving Pipeline 分析（Serving Pipeline Analysis，xllm-npu-pipeline-analysis）

分析已有 profile 的阶段、layer、rank 偏斜和 decode step 气泡。

  • 适用场景： stage layer rank and decode-bubble analysis。
  • 不适用场景： trace collection; generic five-table profiling; raw performance execution。
  • 输出： pipeline phase map; rank-skew analysis; gap and bubble findings。
  • Skill： Serving Pipeline Analysis

昇腾 NPU Profiler（Ascend NPU Profiler，xllm-npu-profiler）

采集昇腾 profile 并生成标准五表分析。

  • 适用场景： Ascend profile collection; generic five-table profile analysis。
  • 不适用场景： fair benchmark conclusion; raw performance run; pipeline-only analysis with complete artifacts。
  • 输出： profile capture; kernel and overlap tables; fusion and efficiency findings。
  • Skill： Ascend NPU Profiler

评测报告生成器（Evaluation Report Writer，xllm-npu-report-writer）

根据已完成 artifacts 和调用方模板生成结构化报告。

  • 适用场景： explicit report rendering from completed artifacts and supplied template。
  • 不适用场景： evaluation execution; fairness policy ownership; template ownership。
  • 输出： structured report; artifact summary; claim-limit section。
  • Skill： Evaluation Report Writer

xLLM 服务管理器（xLLM Service Manager，xllm-npu-server-manager）

启动、验证、smoke、停止并清理 xLLM 服务。

  • 适用场景： explicit xllm service launch ready smoke stop and cleanup。
  • 不适用场景： experiment lifecycle; evaluation workload execution; incident root-cause diagnosis。
  • 输出： launch identity; ready and smoke evidence; cleanup proof。
  • Skill： xLLM Service Manager

性能优化闭环（Performance Optimization Loop，xllm-npu-sota-loop）

执行基于证据的多轮 xLLM 性能优化任务。

  • 适用场景： open-ended measurable performance optimization with code iteration。
  • 不适用场景： one benchmark review; one evaluation run; fixed profiling analysis; incident-only reproduction。
  • 输出： bottleneck budget; candidate ranking; reviewable patches; validated before/after evidence。
  • Skill： Performance Optimization Loop

Triton-Ascend 算子迁移（Triton-Ascend Operator Migration，xllm-npu-triton-migration）

将 Triton-Ascend kernel 迁移为经过验证的 AOT 算子。

  • 适用场景： Triton-Ascend source operator migration to AOT。
  • 不适用场景： runtime integration of an existing operator; non-Triton migration。
  • 输出： Triton kernel; AOT artifact; accuracy and integration evidence。
  • Skill： Triton-Ascend Operator Migration

xllm_ops Runtime 接入（xllm_ops Runtime Integration，xllm-npu-xllm-ops-integration）

把已有 xllm_ops 算子接入 xLLM runtime callsite。

  • 适用场景： integration of an existing xllm_ops operator into xllm runtime。
  • 不适用场景： source operator migration; generic runtime feature。
  • 输出： runtime wrapper; operation registration; build and end-to-end evidence。
  • Skill： xllm_ops Runtime Integration

内部 / 仅显式调用

远程 SSH 执行（Remote SSH Execution，ssh-remote-exec）

为委托流程执行显式请求的 SSH 与远程容器命令。

该能力不能参与隐式主路由。

支持边界

完整支持： xLLM on Ascend 是完整主工作流。 实验性 adapter / 仅 artifact 分析： vLLM-Ascend 与 SGLang 仅限上表列出的 catalog scope。 内部 / 仅显式调用： internal skills 不参与隐式主路由。