tile-ai

TileLang Ascend API 使用最佳实践。提供内存分配、数据搬运、矩阵计算、归约、元素级运算、同步、调度原语等 API 的正确用法和最佳实践。触发：使用 TileLang API 编写 Ascend NPU kernel 时或遇到 API 相关问题时。

TileLang 算子性能调优与潜在性能劣化模式检查。提供性能数据采集、瓶颈诊断、优化实施、效果验证能力；也用于生成或评审算子时对照常见性能劣化模式示例检查当前 kernel 代码。触发：算子精度通过后需要优化性能、性能不及预期时。

Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, update or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or…

根据算子需求生成 TileLang-Ascend 算子设计文档（design.md）。涵盖编程模式选型（Developer/Expert/混合）、API 映射、内存层级规划、Tiling 策略、循环结构、同步策略、验证方案等。触发：设计算子、生成 design.md、算子方案设计、新算子开发、算子实现方案。

基于设计文档生成 TileLang-Ascend 算子实现代码与测试。从 design.md 中提取关键信息，结合 examples/ 中的参考实现生成可运行代码。触发：实现算子、写 kernel、生成代码、算子编码、根据设计文档实现。

TileLang-Ascend 新增 Ascend 专属 T.tile.xxx 小 API 的端到端开发流程。用户要求新增、封装、暴露、实现或测试 ascend_tile.py 中的 T.tile API / Ascend tile primitive 时必须使用本 skill，尤其适用于需要同时打通 Python 前端、C++ lowering/codegen、Ascend C helper、文档和 CI 测试的任务。

TileLang Ascend Developer/Expert 模式选择与 pass_configs 配置指南。当需要确定编程模式、配置 pass_configs、或在两种模式之间转换时触发。API 详情请参考 tilelang-api-best-practices skill。

检查代码格式是否符合 CI 规则。自动检测并安装缺失工具（ruff、clang-format），先运行检查生成报告，使用醒目方式询问用户后，仅在用户同意时执行修复。工作流程：检测环境→自动安装缺失工具→运行检查→生成报告→醒目询问→用户确认→执行修复。使用此技能当：用户要求"格式检查"、"格式化代码"、"代码格式化"、"检查代码格式"、"代码 review"、"代码审查"、"修复格式"、"fix format"、"lint code"、"检查代码规范"、提交 PR 前验证、或需要检查 Python/C++…

对 Stage 1 产出的算子设计文档（DESIGN.md）进行风险优先的检视，生成 REVIEW.md，必须给出明确结论（通过/不通过）与不通过时的具体修改建议。触发：review 设计文档、检视设计、生成 review.md。

根据算子需求生成 TileLang-NPUIR 算子设计文档（DESIGN.md）。涵盖编程模式选型（Developer/Expert/混合）、API 映射、内存层级规划、Tiling 策略、循环结构、同步策略、验证方案等。触发：设计算子、生成 DESIGN.md、算子方案设计、新算子开发、算子实现方案。

根据冻结的 DESIGN.md 生成算子实现（{op}.py：kernel + golden），执行测试并返回三态判定。触发：实现算子、生成 kernel、算子开发、跑精度。

对精度通过的算子实现进行性能调优，产出 {op}.py 与调优日志。触发：性能调优、optimize、性能优化、perf tuning。

TileLang npuir 分支 GitHub 工作流技能。用户提及 commit、push、PR、rebase、upstream、issue、GitHub Actions、gh CLI、分支同步时必须使用本技能。默认遵循 npuir 分支协作规范并提示 Issue 标题使用 [AscendNPU-IR] 或 [npuir] 前缀。

TileLang npuir Vector 算子开发指南。用户提及逐元素、激活函数、归约、广播、sigmoid、rmsnorm、softmax 子流程、vadd/vmul/vexp/vcast/vbrc、向量精度或向量性能优化时必须使用本技能。默认输出必须优先采用 v 前缀 API，而非 npuir_xxx 形式。

TileLang npuir Cube 算子开发指南。用户提及 GEMM、matmul、batch gemm、L1/L0C、load_nd2nz、store_fixpipe、NZ 格式、Cube scope、矩阵分块与流水优化时必须使用本技能。

TileLang npuir 调试辅助技能。用户提及调试 npuir kernel、GDB 附加、IR dump、精度异常定位、编译失败定位、pass 阶段定位、T.print 调试、最小复现缩减时必须使用本技能。

Generate or re-align one `src/tileops/manifest/` entry from a reference-API docs URL. Caller provides the manifest key (`op_name`); skill writes that one entry. Idempotent.

Scaffold a new T2 (L1-direct) Op file from a single `src/tileops/manifest/` entry by following the 7-step playbook in docs/design/ops-design.md. Emits the 17 scaffold slots (S1-S7, S12-S21); leaves family-specific protocol variables, optional hooks, and…

Write tests for the target spec using PyTorch as ground truth, verify they fail on current code.

Drive the full migration for an op family — audit, delegate per-op alignment to align-op, run cross-op cleanup. Two terminal outcomes: SUCCESS opens a PR; CLEANUP_REGRESSION exits blocked without a PR when post-cleanup tests fail.

Per-op orchestrator that brings a single op into alignment with its manifest entry. Classifies the op into one of three cases (green field / interface redesign / minor delta), dispatches to the right path (scaffold-op for new, archive+rescaffold+port for…

Compare each op's code signature against its manifest spec, classify gaps, produce a structured report.

Patch one missing structural field (kernel_map, static_dims) on an existing src/tileops/manifest/ entry. Auto-detects the field via the validator or takes `--field=<name>`. Reference-derivable fields (signature.*, shape_rules, dtype_combos, roofline) belong…

Modify op code to match the manifest-declared interface, making spec tests pass.

Use when working with TileLang layouts — the Layout/Fragment classes, layout inference and its cost model, the CuTe layout algebra and TileLang-to-CuTe conversion, swizzles, or when debugging layout-related passes and their annotations.

Integrate or review a TileLang target backend, target-backend variant, or shared execution backend. Covers language dialects, target and BackendContext contributions, PassPipeline, host/device codegen, execution compatibility, native CMake and packaging,…

Repository-specific build, rebuild, install, and test instructions for tilelang. Use when working in the tilelang repository and the correct commands are needed for building from source, reinstalling after changes, or running project tests.

Use when editing TileLang C++ passes or TVM TIRX code that handles ObjectRef/NodeRef types such as For, Buffer, Var, SBlock, Stmt, PrimExpr, or their *Node raw node counterparts; especially when choosing function parameters, optional values, identity…

Use when editing, reviewing, or documenting TileLang C++ code style, including naming conventions, clang-format choices, headers, TVM ObjectRef/ObjectNode style, FFI-visible fields, and incremental style cleanup plans.

Use when editing TileLang C++ passes or TVM TIRX code that handles ObjectRef/NodeRef types such as For, Buffer, Var, SBlock, Stmt, PrimExpr, or their *Node raw node counterparts; especially when choosing function parameters, optional values, identity…