alibaba
- 164 skills
- 0 followers
- 2.2k repo stars
- yesterday last updated
- ▌ Validate · alibabaEnforce XQUIC test coverage and local validation gates. Use when code, build, or test changes need verification; before commit or pull request; or when asked to build, test, validate, or verify XQUIC. Requires paired happy-path and abnormal-path unit tests, the complete local unit suite, and explicit endpoint coverage or gap reporting for production behavior changes.
- ▌ Gh Fix CI · alibabaDebug and fix failing GitHub PR checks for xquic. Use when a PR, branch, or commit has failed CI, GitHub Actions, build checks, test checks, or requested validation failures that should be diagnosed from logs before editing.
- ▌ Issue Check · alibabaVerify whether a proposed XQUIC bug issue is real by checking the cited RFC or Internet-Draft mechanism against the repository's actual source behavior, tracing any root cause to current file and line evidence, and returning a fail-closed boolean gate. Use when validating an issue draft, investigating a claimed QUIC, HTTP/3, QPACK, or MoQT protocol violation, or before issue-submit may publish an issue.
- ▌ Xquic Build · alibabaPrepare and run XQUIC local compilation. Use when asked to compile, build, configure, or diagnose the local build environment for XQUIC, especially across macOS/Linux. Detect intent and platform, then delegate actual compilation to scripts/validate.sh instead of hand-writing CMake commands.
- ▌ Issue Submit · alibabaDraft and submit evidence-backed XQUIC GitHub bug issues in the repository's CONTRIBUTING and issue-template format. Use when preparing or opening an alibaba/xquic issue that must identify an exact RFC or draft mechanism, expected-versus-actual behavior, reproduction evidence, and optional verified source root cause; requires issue-check to return true before any GitHub submission.
- ▌ Harness Review · alibabaReview XQUIC harness structure for bloat, duplicated facts, trigger overlap, reference drift, and missing hard checks. Use when changing AGENTS.md, harness specs, pipelines, skills, setup scripts, validation routing, or when asked to audit harness design.
- ▌ Xquic Safe Push · alibabaSafely stage, commit, and push XQUIC changes after scoped Git checks. Use when asked to stage, commit, push, publish a branch to a fork, or verify that only intended files will be sent to a remote.
- ▌ Gh Address Comments · alibabaAddress actionable GitHub pull request review comments for xquic. Use when asked to handle PR review feedback, requested changes, unresolved inline threads, or reviewer comments while preserving unrelated edits.
- ▌ Xquic Pr Formatting · alibabaFormat, create, and update concise XQUIC pull request bodies and review-state transitions. Use when asked to prepare or edit an alibaba/xquic PR body, create a draft PR from an already-pushed branch, update a PR after code changes pushed a new head, or move a validated draft toward review.
- ▌ Xquic Pr Pre Review · alibabaFail closed on published draft XQUIC code pull requests before ready-for-review state. Use when asked to pre-review a code PR or move it out of draft; verify current base-to-head code, protocol conformance, validation coverage, adversarial bad cases, hot-path performance, and memory safety in a five-part local report.
- ▌ Classify Issue Complexity · alibabaClassify XQUIC GitHub issues with an evidence-backed L1-L4 resolution-complexity label while keeping validity, priority, severity, and disposition separate and skipping uniquely labeled issues unless the user explicitly forces a recheck. Use when triaging one issue or a batch of issues, creating an issue-complexity backlog, or applying or correcting L1-L4 GitHub labels.
- ▌ Flexlb Development · alibaba bundleFlexLB 开发、缺陷修复、重构和代码审核规范。适用于 RTP-LLM 的 rtp_llm/flexlb 及其调用方,重点约束类流程说明、成员组织、请求与资源生命周期、修改范围、验证及提交;不用于无关的模型推理代码或线上运维。
- ▌ Workflow First Design · alibaba bundle复杂功能、框架和跨模块重构的整体设计与落地规范。用于确定业务流程、模块职责、跨组件状态与资源协议、失败协调、并发性能取舍、迁移清理和集成验收;与 evidence-first-implementation 配合落实具体类和方法。局部实现问题不强制升级为架构设计。
- ▌ Evidence First Implementation · alibaba具体类、方法与局部组件的实现规范,用于功能开发、缺陷修复和重构。依据真实代码和行为契约,落实类职责、字段归属、接口语义、封装与继承、资源清理、局部并发和测试,清理被替代的实现;跨组件设计与 workflow-first-design 衔接,简单修改直接实施。
- ▌ Open Code Review Delegate · alibabaDelegation mode for open-code-review (OCR). Instead of OCR calling an LLM endpoint, this skill instructs the host agent to perform the code review itself, using OCR only for deterministic engineering: file selection and rule resolution. Use when the host agent should drive the review with its own LLM capabilities.
- ▌ Bmad Spec 2 · alibaba bundleDistill any intent input into the SPEC kernel + companions — the canonical, preservation-validated machine contract for downstream work. Use when the user says "create a spec", "distill this into a spec", "validate this spec", or "update the spec".
- ▌ Bmad Dev Auto 2 · alibaba bundleOne iteration of an unattended development loop. Use when invoked by name.
- ▌ Bmad Quick Dev 2 · alibaba bundleImplements any user intent, requirement, story, bug fix or change request by producing clean working code artifacts that follow the project's existing architecture, patterns and conventions. Use when the user wants to build, fix, tweak, refactor, add or modify any code, component or feature.
- ▌ Bmad Code Review 2 · alibaba bundleReview code changes adversarially using parallel review layers (Blind Hunter, Edge Case Hunter, Acceptance Auditor) with structured triage into actionable categories. Use when the user says "run code review" or "review this code"
- ▌ Bmad Agent Builder 2 · alibaba bundleBuilds, edits or analyzes Agent Skills through conversational discovery. Use when the user requests to "Create an Agent", "Analyze an Agent" or "Edit an Agent".
- ▌ Bmad Workflow Builder 2 · alibaba bundleBuilds, converts, and analyzes workflows and skills. Use when the user requests to "build a workflow", "modify a workflow", "quality check workflow", "analyze skill", or "convert a skill".
- ▌ Object Object 2 · alibaba bundleSets up {module-name} module in a project. Use when the user requests to 'install {module-code} module', 'configure {module-name}', or 'setup {module-name}'.
- ▌ Bmad Spec · alibaba bundleDistill any intent input into the SPEC kernel + companions — the canonical, preservation-validated machine contract for downstream work. Use when the user says "create a spec", "distill this into a spec", "validate this spec", or "update the spec".
- ▌ Bmad Dev Auto · alibaba bundleOne iteration of an unattended development loop. Use when invoked by name.
- ▌ Bmad Quick Dev · alibaba bundleImplements any user intent, requirement, story, bug fix or change request by producing clean working code artifacts that follow the project's existing architecture, patterns and conventions. Use when the user wants to build, fix, tweak, refactor, add or modify any code, component or feature.
- ▌ Bmad Code Review · alibaba bundleReview code changes adversarially using parallel review layers (Blind Hunter, Edge Case Hunter, Acceptance Auditor) with structured triage into actionable categories. Use when the user says "run code review" or "review this code"
- ▌ Loongsuite Pilot Ops 2 · alibaba bundle安装、卸载 LoongSuite Pilot(面向研发团队的 AI Coding 数据采集工具, 支持 Cursor / Qoder / Claude Code / Codex)。当用户提到 loongsuite-pilot 安装、 pilot 部署、Node.js 前置条件、nvm 配置、Linux 7U glibc 补丁、SSH 远程安装、 新增或追加 SLS flusher、自定义 SLS/WebTracking 上报配置时, 务必使用本 Skill——即使用户只问“怎么装”也应触发。安装完成后可在 ~/.loongsuite-pilot/skills/loongsuite-pilot-ops/references/ 目录下找到诊断排查与观测面板的详细参考文档。
- ▌ Auto Develop 2 · alibaba bundleOPC-Starter 智能开发技能。AI 亲和的 React Boilerplate 项目开发规范,支持动态上下文感知、TDD 驱动开发、Agent Studio 扩展。适用于认证系统、组织架构、Agent 工具、数据同步等模块的迭代开发。
- ▌ Object Object · alibaba bundleSets up {module-name} module in a project. Use when the user requests to 'install {module-code} module', 'configure {module-name}', or 'setup {module-name}'.
- ▌
- ▌ Use React Intl Universal · alibaba bundleBest-practice workflow for building and inspecting high-quality internationalized software with react-intl-universal, react-intl-universal-extract, and this skill's rules/scripts. Use it directly for repositories that use react-intl-universal, and use its language-quality, UI-quality, localized-UI inspection, and locale-synchronization principles as reference guidance for other i18n stacks.
- ▌ Umodel Rca · alibaba bundleModel-guided autonomous root-cause analysis over a UModel object-graph semantic layer. Use with `umodel-query` when diagnosing service degradation, SLO breach, latency/error spikes, incidents, outages, or questions like why a service is slow, degraded, erroring, or timing out. The skill guides evidence collection across object graph, metrics, logs, topology, runbooks, deployments, config changes, business traffic, and runtime capacity; it is especially useful for ruling out red herrings such as recent but unrelated deployments. Triggers: root cause, RCA, incident, SLO breach, outage, postmortem, latency spike, error spike, timeout, 根因分析, 故障排查, 告警定位, 为什么慢, SLO 击穿.
- ▌ Umodel Query · alibaba bundleRead from a UModel object-graph semantic layer with the `umctl` CLI (MCP alternative noted). Three kinds of read: (1) entities & relationships / topology (`.entity`, `.topo`) and (2) the model itself (`.umodel`, and `.entity_set` methods, including entity-linked `list_skills`) return real rows; (3) metrics & logs (`get_metrics` / `get_logs`) return an executable *plan* — PromQL / Elasticsearch DSL with the entity id pre-substituted — that you run against the backend. Against a PaaS endpoint the same calls return data rows instead of a plan. Use to query or read UModel entities, relations, topology, or model metadata; to read a service's metrics or logs; to look up services and dependencies; or to discover what objects, datasets, and methods exist. For root-cause analysis on top of these reads, see the `umodel-rca` skill. Triggers: UModel, object graph, .entity / .topo / .umodel / .entity_set, query entities, read topology, read metrics / logs, get_metrics / get_logs, list services / dependencies / datasets /
- ▌ Umodel Skill Runner · alibabaUse when an EntitySet exposes list_skills, a user requests a UModel or RunbookSet Skill, or an entity task supplies SKILL.md. Triggers: list_skills, list_knowledge, RunbookSet Skill, UModel Skill, execute skill, run skill, 执行技能, 运行技能, Runbook 技能.
- ▌
- ▌
- ▌
- ▌
- ▌
- ▌ Sls Dashboard Builder · alibaba bundle当任务需要创建、修改、扩展或重组阿里云 SLS 的 dashboard JSON 或可导入的大盘配置时使用;尤其适用于线上大盘、强对比的分析看板、已校验的查询包,或需要专业中文标签与指标定义的运维向大盘。
- ▌ Loongsuite Pilot Insight · alibaba bundle基于 LoongSuite Pilot / AI Coding Agent 日志生成事件洞察、组织洞察、数据质量、研发效能和 AI Native 使用类 SLS 报表时使用;包含 AI Coding 事件表语义,以及团队报表可选的部门维表、dept_user 组织关系、指标口径和公共 CTE,通常与 sls-dashboard-builder 一起使用。
- ▌
- ▌ PDF Extractor · alibaba bundleExtract text, tables, and form data from PDF documents for analysis and processing. Use when user asks to extract, parse, or analyze PDF files.
- ▌ Sales Analytics · alibabaDatabase schema and business logic for sales data analysis including customers, orders, and revenue.
- ▌
- ▌ Grouped Tools Test · alibabaTest skill for groupedTools. When executing this skill, use the record_result tool to record the result value.
- ▌ Inventory Management · alibabaDatabase schema and business logic for inventory tracking including products, warehouses, and stock levels.
- ▌ Flight Profiler Stack · alibabaInspect stack frames of a running Python process including thread stacks, native C-level frames, and async coroutine stacks. Use this to see what each thread is currently doing and diagnose hangs, deadlocks, or stuck async tasks.
- ▌ Flight Profiler Trace · alibabaTrace execution time of Python method invocations in a live process with call stack visualization. Use this to see a hierarchical call tree showing which sub-calls are slow and where time is spent within a function.
- ▌ Flight Profiler Watch · alibabaDisplay input/output args, return object and cost time of a Python method invocation in a live process. Use this to observe function behavior at runtime — see what arguments go in, what comes back, and how long it takes. This is the most frequently used diagnostic command.
- ▌ Flight Profiler Attach · alibabaPrerequisites and connection details for attaching to a live Python process with PyFlightProfiler. Read this skill first before using any other PyFlightProfiler command (watch, trace, stack, perf, etc.).
- ▌ Flight Profiler Module · alibabaTranslate a file path to its Python module name in a running Python process. Use this when you know a file path but need the module name for commands like watch, trace, or reload.
- ▌ Flight Profiler Reload · alibabaHot-reload a Python function implementation from the latest file content without restarting the process. Use this to apply code fixes live after identifying a bug with watch/trace.
- ▌ Flight Profiler Vmtool · alibabaInspect live Python class instances in a running process — find all instances of a class, examine their attributes/state, or invoke methods on them. Also supports forcing GC. Use this when you need to probe object-level runtime state beyond what watch/getglobal can provide.
- ▌ Flight Profiler Getglobal · alibabaInspect Python module global field or class static field value in a running Python process. Use this for safe, read-only inspection of global variables, configuration objects, or class-level state.
- ▌ Gen Plan · alibaba bundleGenerate a structured implementation plan from an evidence draft. Validate paths, obtain configured independent Codex and Qoder reviews, synthesize available advice against repository evidence, preserve the draft, and produce testable acceptance criteria and validation steps.
- ▌
- ▌ Gpu Kernel Baseline · alibabaLearn the target framework from gpu-wiki and implement a baseline GPU kernel. Use this skill to understand compute semantics, determine the target platform and framework, search reference implementations, and produce a correct V0 baseline with performance records for later profile-driven optimization.
- ▌ Ppu Acu Joint Profile · alibaba bundleChoose and run ACU-only, adaptive PPU in-kernel timeline, or optional bounded joint analysis for a PPU kernel. Use for device-level bottleneck diagnosis, kernel-internal critical-path questions, or evidence that genuinely needs both; do not require all three modes.
- ▌ Gpu Kernel Episode Loop · alibabaRun the evidence loop of one long-horizon GPU kernel optimization episode. Use this skill to reconstruct the incumbent, profile and localize a bottleneck, research progressively, plan one coherent direction, implement and repair, validate development correctness and performance, and record every decisive experiment in the episode journal.
- ▌ Opt Trace Mining · alibaba bundleMine a per-kernel optimization trace — a git repository capturing successive versions of one kernel being optimized — into structured, gate-validated optimization-experience records for the GPU kernel wiki. Use when asked to distill an optimization run, a kernel_opt trace directory or a version ladder into wiki records; to report what such a run actually achieved; to build, extend, re-run or validate the staging store behind those records; or to explain how a trace-derived record's number, snippet or provenance was established.
- ▌ Query Bridge Agent · alibabaQuickly extract a typed GPU-Wiki query intent from prose. Never inspect or query a knowledge store.
- ▌ Session Trace Mining · alibaba bundleMine AI coding-agent session transcripts into structured, gate-validated GPU-kernel optimization records for the wiki. Use when asked to turn vibe-coding sessions, Codex rollout logs, or Claude Code project transcripts into wiki records; to summarise what a kernel-optimization session achieved; to build or extend a session-trace store; or to re-run and validate one. Also use when asked how a session-derived record's number, snippet, or provenance was established.
- ▌ Autonomous Gpu Kernel Timeline · alibaba bundleLet AKA autonomously add, run, inspect, and revise intra-kernel timeline probes for standalone CUDA/inline PTX or CuTe DSL when ordinary benchmark, NSYS, or NCU evidence cannot answer a specific kernel-internal timing question.
- ▌
- ▌
- ▌
- ▌
- ▌ Speckit Plan · alibaba bundleExecute the implementation planning workflow using the plan template to generate design artifacts.
- ▌
- ▌ Speckit Tasks · alibaba bundleGenerate an actionable, dependency-ordered tasks.md for the feature based on available design artifacts.
- ▌ Speckit Specify · alibaba bundleCreate or update the feature specification from a natural language feature description.
- ▌
- ▌
- ▌ Check · alibabaRuns .continue/checks locally against the current diff, simulating a PR review before pushing.
- ▌ Esa Deploy · alibaba bundleDeploys React/Vite applications to Alibaba Cloud ESA Pages. Use when the user needs to deploy frontend applications, manage deployments, configure custom domains, or troubleshoot deployment issues on ESA (Edge Security Acceleration) platform.
- ▌ Auto Develop · alibaba bundleOPC-Starter 智能开发技能。AI 亲和的 React Boilerplate 项目开发规范,支持动态上下文感知、TDD 驱动开发、Agent Studio 扩展。适用于认证系统、组织架构、Agent 工具、数据同步等模块的迭代开发。
- ▌ Test Driven Dev · alibaba bundle[project] OPC-Starter 测试驱动开发规范。管理和执行 Playwright E2E 测试、演示模式测试、单元测试的完整工作流。适用于新增功能时补充测试用例、运行认证/注册/演示测试、维护测试规划文档。当用户提到测试、E2E、Playwright、演示、用例、质量、补充测试时使用。
- ▌ Design System UI · alibaba bundleCreate distinctive, production-grade frontend interfaces with high design quality using Mobile First strategy. Analyzes design systems from screenshots, generates Tailwind CSS + Shadcn UI configurations, and retrofits existing projects. Optimized for Claude Opus 4.5 with structured decision trees and self-verification checkpoints.
- ▌ Ihs Repo Harness · alibaba bundleIHS(IDE Harness Score)仓库健康评估技能。用于衡量 AI Coding IDE 驾驭能力下的代码仓技术债变化趋势,输出 IHS.md 评测报告(代码腐化、测试信号、文档对齐、变好/变坏结论)。
- ▌ Bmad Agent Builder · alibaba bundleBuilds, edits or analyzes Agent Skills through conversational discovery. Use when the user requests to "Create an Agent", "Analyze an Agent" or "Edit an Agent".
- ▌ AI Friendly Audit · alibaba bundle代码仓库 AI 亲和度审计工具,支持前端、后端、全栈等各类项目。此技能用于检查给定的代码仓库对 AI Coding 工具的友好程度,生成详细的分析报告和改进建议。当用户需要评估代码仓库是否适合 AI 辅助开发、希望提升仓库的 AI 可操作性、或准备引入 AI Coding 工具前进行仓库评估时,应使用此技能。支持 TypeScript/JavaScript、Go、Java/Kotlin、Python、Rust 等主流技术栈。融合 OpenAI Harness Engineering 方法论,评估 Outer Loop(反馈闭环、评估门禁、机械化不变量)建设。
- ▌ Bmad Workflow Builder · alibaba bundleBuilds, converts, and analyzes workflows and skills. Use when the user requests to "build a workflow", "modify a workflow", "quality check workflow", "analyze skill", or "convert a skill".
- ▌ Cpu · alibaba bundleMNN CPU 后端(ARM / x86_64 / RISC-V 三侧)的总入口,只做分流不承载技术内容。下分 `optimize/`(为什么慢、该改哪一层)与 `kernel/`(这条 kernel 怎么写对、怎么被选中)两个分支,`shared/` 放两者共用的构建测试跑分命令、env 开关注册表与 RISC-V 开发板远端验证纪律。做 CPU 侧的工作但还不确定该进哪个分支,或需要三侧结构差异对照(第二张函数表按什么分、二级表怎么构造、`Precision_Low` 语义、ISA A/B 怎么做——「三侧不同构对照表」全树唯一一份,就在本文件)时读这里。
- ▌ Hexagon Optimization · alibaba bundleMNN Hexagon/HVX/HMX DSP 后端(`source/backend/hexagon`)的优化、重构、构建与回归验证。覆盖设备实测的相位分解、测量纪律、v79/v81 双架构差异、常见瓶颈模式、已否证方向与 cDSP crash 诊断。
- ▌ Test CI · alibaba bundleRun MNN tests / benchmarks on host or real devices. Covers two parallel tracks — (1) the regression / CI suite (static checks, host-side tests, on-device Android arm64 matrix via ./test.sh + test_stages.json) and (2) one-command iOS real-device LLM benchmarking (prefill/decode tok/s, branch comparison). Use when the user asks to run the tests, run CI, smoke-test a build, verify a change on a device, benchmark on-device (Android or iPhone/iPad), or add / select / retune a test stage.
- ▌ Qnn Debug · alibaba bundleMNN QNN(高通 HTP/NPU)后端的问题定位、修复与算子适配。覆盖运行时报错(1002/6000/1003/6004)、模型转换/算子校验失败、推理结果异常、精度不达标、以及性能问题;并指导为 QNN 新增/适配算子。核心方法:用 MNN2QNNModel --dump_intermediate_outputs 一次性导出 QNN 全部中间张量,与 CPU(fwd=0) 基线逐张量对比,直接定位首个出错算子(旧的 testMNNFromOnnx.py 截断二分已降为回退手段);区分“真 bug/量化/HTP fp16 精度”;新增算子先查 SDK 算子文档(MasterOpDef/HtpOpDefSupplement)再实现。
- ▌ Add New Op · alibaba bundle为 MNN 框架新增算子。包含 Schema 定义、形状计算、几何计算、后端实现、单元测试的完整 TDD 流程。分 5 步执行,每步有独立测试标准。
- ▌ Cpu Kernel · alibaba bundleMNN CPU 后端 kernel 开发分支(`skills/cpu/` 下,另一分支是 `cpu/optimize` 性能归因)。覆盖标量 oracle → C++ SIMD → intrinsic → 汇编的四级实现阶梯、pack/kernel ABI 契约(tile、cell stride、后处理参数)、CoreFunctions 派发表注册与二级表安全构造、跨 ISA × 精度的正确性门禁,以及 AArch64 / x86_64 / RISC-V 三份实现参考。为 CPU 后端新写或移植 kernel(NEON / SSE / AVX / RVV intrinsic 或 .S 汇编)、新增一层 ISA、改 pack mode 或 tile 参数、把 kernel 挂进函数表时使用。已定位到 kernel 才进本分支。
- ▌ Cpu Optimize · alibaba bundleMNN CPU 后端性能归因分支(`skills/cpu/` 下,另一分支是 `cpu/kernel` kernel 开发)。按五层(Runtime 线程 / Executor 调度 / Layout 内存 / Dispatch 函数表 / Kernel ISA)定位瓶颈,含 bound 类型判定、op 级实验回路、跨框架 op 对 op 对比、跨层不一致的事后定位,以及 ARM / x86_64 / RISC-V 三侧「我到底跑在哪条 ISA 路径上」的诊断面。CPU 上算子或 LLM prefill/decode 慢、线程数或内存占用异常、出现性能回归、要与外部推理框架逐算子对比时使用。
- ▌ Bugfix · alibaba bundleMNN 各类正确性/回归 bug 的排查入口,按 bug 类别分册组织,本文件只做症状分流。分册:内存别名与生命周期(arena reuse、`MemChunk`、融合引入的别名竞争)、量化误差与导出侧权重损坏(低 bit 打包、导出分块、PyTorch MPS/CUDA 大张量静默错误)、host 侧并发/线程竞争(共享所有权的引用计数被写坏、析构链崩溃、TSAN A/B 与编译期哨兵)、fp16 表示能力不足(长序列复读、position 塌缩,以及「实时计算→预计算查表」重构的三类陷阱)、GPU shader 越界与 command buffer 故障、后端 kernel 隐式假设违反(causal mask、layout 约定)、持久化缓存误信(weight-mmap sync 自我污染、跨模型缓存复用)、逐 run 不同的非确定性(未初始化内存/堆垃圾依赖、多线程动态分发×异构 kernel)。用户报告 MNN 输出乱码/退化、单测或 golden 对不上、改动后回归、换后端结果不同、开某开关才错、结果每次跑都不一样、或崩在析构链上且只在后台线程异步释放时偶现时使用。
- ▌ Retrospective · alibaba任务完成后的反思与经验沉淀。仅在任务非平凡且出现调试、失败修复、反复试错、明显误判或可复用教训时使用;简单执行、查询、常规 CI 通过等无新增经验的任务可跳过。
- ▌ Metal Optimize · alibaba bundleMNN Metal 后端 op/kernel 开发与优化入口。索引各份 sub-doc:性能问题诊断流程(op 单测 + 对手基准定位真瓶颈,优化任务第一站)、kernel 开发规范与优化知识库(命名/写法/GEMV/GEMM/attention + 手段方法论)、算子融合全链路(导出图→converter→Metal 单 dispatch + 融合方法论)、运行时调度(fence/content-cache/H2D/replay + 调度方法论)、构建测试基线、env 开关注册表。根据当前任务选择性阅读对应 sub-doc。
- ▌ Opencl Optimize · alibaba bundleMNN OpenCL 后端 kernel 性能优化与新特性集成。覆盖 benchmark 基线、kernel 优化迭代、集成验证全流程,以及新 OpenCL 特性适配、.cl + codegen 双轨、kernel 选路、packed weight 设计、tune 机制、Android 真机验证等参考知识。
- ▌ Support New LLM · alibaba bundle为 MNN 框架添加新的 LLM 模型支持。支持从 HuggingFace/ModelScope 下载模型,分析架构,添加映射,Hook 对齐测试,导出 MNN 模型。采用 TDD 模式,分 6 步执行,每步有独立测试标准。
- ▌ Vulkan Optimize · alibaba bundleMNN Vulkan 后端 kernel/算子性能优化与新特性集成。覆盖 benchmark 基线、kernel 优化迭代、集成验证全流程,以及 GLSL .comp + makeshader 双轨、conv1x1/attention dispatcher 多路径(coopMat/subgroup/nosubgroup)、cooperative matrix、packed weight 设计、pipeline cache、CPU 侧调度瓶颈、Adreno/Mali/Apple 多 vendor 真机验证等参考知识。
- ▌ E2e · alibaba bundleLoongCollector E2E 测试全流程指南:设计、编写、运行和调试。当需要编写新 E2E 测试、运行现有测试、或排查 E2E 测试失败时使用此 skill。
- ▌
- ▌
- ▌