Figma Design System Builder
warpdotdev/warp
Builds or updates a design system in Figma from a codebase in ordered phases: discovery, variables and tokens, components, theming and documentation, with checkpoints.
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。
Install Tilelang Op Develop
$ npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a claude-codeProject install by default; add -g for ~/.claude/skills/.
$ gh skill install tile-ai/tilelang-ascend tilelang-op-develop --agent claude-codeProject scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).
$ git clone --depth 1 https://github.com/tile-ai/tilelang-ascend.git skills-src && mkdir -p .claude/skills && cp -r skills-src/.agents/skills/tilelang-op-develop .claude/skills/tilelang-op-develop && rm -rf skills-srcUse ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.
Claude Code skills documentation · loads skills from .claude/skills/
Install the "tilelang-op-develop" agent skill from https://github.com/tile-ai/tilelang-ascend/tree/ascendc_pto/.agents/skills/tilelang-op-develop into .claude/skills/tilelang-op-develop/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "tilelang-op-develop", then confirm the skill loads.Claude Code copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$skill-installer install https://github.com/tile-ai/tilelang-ascend/tree/ascendc_pto/.agents/skills/tilelang-op-developType this inside Codex. $skill-installer <name> installs a curated skill from openai/skills. The installer writes to $CODEX_HOME/skills (default ~/.codex/skills). Restart Codex if the skill does not show up.
$ npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a codexProject install goes to .agents/skills/; add -g for ~/.codex/skills/.
$ gh skill install tile-ai/tilelang-ascend tilelang-op-develop --agent codexProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/tile-ai/tilelang-ascend.git skills-src && mkdir -p .agents/skills && cp -r skills-src/.agents/skills/tilelang-op-develop .agents/skills/tilelang-op-develop && rm -rf skills-srcUse ~/.agents/skills/ instead of .agents/skills for a personal install.
Codex skills documentation · loads skills from .agents/skills/
Install the "tilelang-op-develop" agent skill from https://github.com/tile-ai/tilelang-ascend/tree/ascendc_pto/.agents/skills/tilelang-op-develop into .agents/skills/tilelang-op-develop/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "tilelang-op-develop", then confirm the skill loads.Codex copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a cursorProject install goes to .agents/skills/; add -g for ~/.cursor/skills/.
$ gh skill install tile-ai/tilelang-ascend tilelang-op-develop --agent cursorProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/tile-ai/tilelang-ascend.git skills-src && mkdir -p .cursor/skills && cp -r skills-src/.agents/skills/tilelang-op-develop .cursor/skills/tilelang-op-develop && rm -rf skills-srcUse ~/.cursor/skills/ instead of .cursor/skills for a personal install.
Cursor skills documentation · loads skills from .cursor/skills/, .agents/skills/, .claude/skills/, .codex/skills/
Install the "tilelang-op-develop" agent skill from https://github.com/tile-ai/tilelang-ascend/tree/ascendc_pto/.agents/skills/tilelang-op-develop into .cursor/skills/tilelang-op-develop/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "tilelang-op-develop", then confirm the skill loads.Cursor copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gemini skills install https://github.com/tile-ai/tilelang-ascend.git --path .agents/skills/tilelang-op-develop--scope user (default) or --scope workspace; --path is the subfolder of the repo that holds the skill; --consent skips the security confirmation prompt.
$ npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a gemini-cliProject install goes to .agents/skills/; add -g for ~/.gemini/skills/.
$ gh skill install tile-ai/tilelang-ascend tilelang-op-develop --agent gemini-cliProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/tile-ai/tilelang-ascend.git skills-src && mkdir -p .gemini/skills && cp -r skills-src/.agents/skills/tilelang-op-develop .gemini/skills/tilelang-op-develop && rm -rf skills-srcUse ~/.gemini/skills/ instead of .gemini/skills for a personal install, then run /skills reload.
Gemini CLI skills documentation · loads skills from .gemini/skills/, .agents/skills/
Install the "tilelang-op-develop" agent skill from https://github.com/tile-ai/tilelang-ascend/tree/ascendc_pto/.agents/skills/tilelang-op-develop into .gemini/skills/tilelang-op-develop/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "tilelang-op-develop", then confirm the skill loads.Gemini CLI copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gh skill install tile-ai/tilelang-ascend tilelang-op-developInstalls for Copilot at project scope by default; add --scope user for a personal install. Preview a skill first with gh skill preview. Needs GitHub CLI 2.90.0 or later (public preview).
$ npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a github-copilotProject install goes to .agents/skills/; add -g for ~/.copilot/skills/.
$ git clone --depth 1 https://github.com/tile-ai/tilelang-ascend.git skills-src && mkdir -p .github/skills && cp -r skills-src/.agents/skills/tilelang-op-develop .github/skills/tilelang-op-develop && rm -rf skills-srcUse ~/.copilot/skills/ instead of .github/skills for a personal install. Commit .github/skills so cloud agent and code review can use it.
GitHub Copilot skills documentation · loads skills from .github/skills/, .claude/skills/, .agents/skills/
Install the "tilelang-op-develop" agent skill from https://github.com/tile-ai/tilelang-ascend/tree/ascendc_pto/.agents/skills/tilelang-op-develop into .github/skills/tilelang-op-develop/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "tilelang-op-develop", then confirm the skill loads.GitHub Copilot copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a opencodeOpenCode documents no install command of its own. Project install goes to .agents/skills/; add -g for ~/.config/opencode/skills/.
$ gh skill install tile-ai/tilelang-ascend tilelang-op-develop --agent opencodeProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/tile-ai/tilelang-ascend.git skills-src && mkdir -p .opencode/skills && cp -r skills-src/.agents/skills/tilelang-op-develop .opencode/skills/tilelang-op-develop && rm -rf skills-srcUse ~/.config/opencode/skills/ instead of .opencode/skills for a personal install.
OpenCode skills documentation · loads skills from .opencode/skills/, .claude/skills/, .agents/skills/
Install the "tilelang-op-develop" agent skill from https://github.com/tile-ai/tilelang-ascend/tree/ascendc_pto/.agents/skills/tilelang-op-develop into .opencode/skills/tilelang-op-develop/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "tilelang-op-develop", then confirm the skill loads.OpenCode copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
Facts
- Skill name
tilelang-op-develop- GitHub stars
- 400
- Token cost
- ~2.6k tokens
- SKILL.md length
- 812 words
- Files
- 9 (incl. references)
- Skills in repo
- 20
- Repo updated
- First seen
- Licence
- MIT
At a glance
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。
- Works in 4 steps: 从 design.md 中提取的信息（只取这些） → 参考来源（优先级高于 design.md 伪代码） → 代码生成流程 → …
- Tasks that involve Design tokens
- SKILL.md covers 1. 从 design.md 中提取的信息（只取这些）, 2. 参考来源（优先级高于 design.md 伪代码）, 3. 代码生成流程 and 4. Skill 反馈采集, plus 1 more section
- Calls ruff and python
What it does
Tilelang Op Develop is an agent skill from tile-ai/tilelang-ascend. 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。
Its SKILL.md is about 2.6k tokens, which your agent loads only when the skill is triggered. The skill folder holds 10 other files, including reference files (for example `examples/code-skeleton.md`, `references/ascend-constraints.md` and `references/checklist.md`).
It sits in Frontend & Design, covering Design tokens. The repository describes itself as: Ascend TileLang adapter. The licence is MIT.
When your agent uses it
- Tasks that involve Design tokens
Example prompts
- “/tilelang-op-develop”
Requirements
- Python 3
Workflow steps
4 steps, taken from the step headings in SKILL.md.
What it can do on your machine
Read from SKILL.md and the folder at commit 3d17c28. It shows what the files ask for, not the result of running them.
Tool permissions
Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.
From allowed-tools in the SKILL.md frontmatter.
Runs code
Shell commands in SKILL.md call:
ruffpython
From the folder's file list and the shell code blocks in SKILL.md.
Network
No URLs in SKILL.md.
From URLs in SKILL.md, links to its own repository left out.
Credentials
Names no API keys, tokens, secrets or passwords.
From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.
Context cost
Tilelang Op Develop loads about 2.6k tokens when it runs, and up to ~16k if it reads all its reference files. Until then it costs about 35 tokens; SKILL.md has 812 words of instructions outside code blocks.
Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.
Safety
Auto-check passed
The automated check found no risky patterns in SKILL.md.
Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.
SKILL.md
The full file from tile-ai/tilelang-ascend at commit 3d17c28, republished under its MIT licence (© tile-ai). 812 words, ~2,645 tokens.
Download SKILL.mdSave it as
.claude/skills/tilelang-op-develop/SKILL.md (or your agent's skills folder). This skill also uses 8 other files; get the full folder from GitHub.- name
- tilelang-op-develop
- description
- 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 算子代码生成
基于设计文档（design.md）和已有示例，生成可运行的算子实现与测试。
1. 从 design.md 中提取的信息（只取这些）
design.md 可能很长，只提取以下字段，忽略其余内容：
| 提取字段 | 所在章节 | 用途 |
|---|---|---|
| 数学公式 | §1 概述 | 理解计算逻辑 |
| 算法步骤分解 | §1 算法描述 | 确定计算顺序 |
| API 映射表 | §3 API 映射设计 | 核心：每步用哪个 TileLang API |
| 伪代码 | §3 计算伪代码 | 核心：代码骨架 |
| 输入输出 shape 和 dtype | §4 数据规格 | 函数签名和测试数据 |
| block 大小 | §5 Tiling 策略 | 分块参数 |
| pass_configs | §7 同步策略 | JIT 配置 |
| Golden 函数 | §9.1 Golden 函数 | 测试对比基准 |
| 测试用例表 | §9.2 L0 门槛测试计划 | 测试配置 |
| 精度标准 | §9.3 精度标准 | 混合容差：atol / rtol / max_abs_error_limit / required_matched_ratio（按 dtype） |
| 路径性能可行性表 | §5/§6 | GM pass、DMA transaction、GM 标量访问、地址计算和并行度 |
| 性能可行性哨兵 | §9 | 每条路径最坏 dtype/最大任务数 case 与单 case 超时预算 |
明确忽略的内容（这些容易误导）：
- 模式选型的分析推理过程
- 内存预算的计算过程和多轮优化迭代
- 仅忽略没有量化证据的笼统风险；凡是包含具体 shape、dtype、超时、GM/DMA 成本或回退路径的风险必须提取并作为验收约束
- 交付清单（仅是文件列表）
- 任何标注为"待确认"的内容
2. 参考来源（优先级高于 design.md 伪代码）
当 design.md 伪代码与 examples/ 中同类实现有冲突时，以 examples/ 为准。
2.1 API 用法和模式选择
- API 用法：查阅 tilelang-api-best-practices SKILL.md 及其 references 目录
- 编程模式和 pass_configs：查阅 tilelang-programming-model-guide SKILL.md 及其 references 目录
2.2 同类算子示例
生成代码前，必须查阅 examples/ 中的同类算子：
| 算子类型 | 参考示例 |
|---|---|
| 逐元素运算（add/mul/sigmoid/relu） | examples/elementwise/、examples/activation/ |
| 归约运算（reduce_sum/max/min） | examples/reduce/ |
| 归一化（softmax/layernorm/rmsnorm） | examples/softmax/、examples/normalization/ |
| GEMM | examples/gemm/、examples/developer_mode/gemm_developer.py |
| 融合算子 | examples/flash_attention/、examples/pipeline/、examples/developer_mode/matmul_add_developer.py |
| Developer 模式 | examples/developer_mode/ |
| transpose / layout transform | examples/transpose/transpose.py（提取结构谓词、 |
| 连续 suffix-record 聚合搬运和通用 fallback；不得照抄具体 perm/shape 分支） |
查阅示例时关注：
- Kernel 结构：
T.Kernel参数、cid/vid用法 - Buffer 分配方式：shape 和 dtype
- pass_configs 配置：该类算子实际使用哪些开关
- 数据搬运：
T.copy的索引写法 - CV 交互（融合算子，按模式）：Developer 默认
threads=2+ 片上直连（无 workspace_idx）；Expert/混合或回退才看 workspace_idx、数量、shape
3. 代码生成流程
开始生成代码前，必须先读取并遵从 references/ascend-constraints.md——其中定义了两条强制规则：
- §1 算子 kernel 划分原则：支持域完整 + fallback 审计，未覆盖输入立即返回
[DESIGN_ERROR] - §2 Host 侧 Buffer 操作约束：算子核心逻辑全部在 kernel 内实现，host 侧禁止改数据指针 / 真实重排 / 改写 buffer / 用新 buffer 作弊 / 隐式触发 aclnn 调用
以下流程步骤均建立在这些约束之上；生成代码后会在步骤 5 上库前检查中逐项复核。
步骤 1：读取设计文档
读取 design.md，按 §1 的表格提取字段。
步骤 2：查找参考示例
在 examples/ 中找到最相似的算子实现，完整阅读其代码并记录技术决策：
必须记录的技术决策（从参考实现中提取）：
| 决策项 | 示例值 | 说明 |
|---|---|---|
| 内存层级 API | alloc_L1/L0C/ub（显式）或 alloc_shared/fragment（自动） | 决定内存分配方式 |
| 同步策略 | 手动 barrier_all/set_flag 或自动同步 | 决定同步代码 |
| pass_configs | AUTO_SYNC: True，融合算子需 AUTO_CV_COMBINE: True + AUTO_CV_SYNC: True | 决定 JIT 配置 |
| 核分离方式 | T.Scope("C"/"V") 或无显式分离 | 决定核间协作方式 |
| CV 交互（融合算子，按模式） | Developer：threads=2 + 单 cid 轴 + 片上直连（无 workspace_idx）；Expert/混合/回退：{数量: 3, shape: [block_num, block_M, block_N], idx: [4,5,6]} | Developer 默认消除 workspace/vid，见 mode-examples.md §6 |
对比差异分析（如有 design.md）：
| 项目 | design.md 方案 | 参考实现方案 | 选择理由 |
|---|---|---|---|
| 内存层级 API | |||
| 同步策略 | |||
| pass_configs | |||
| CV 交互 ⭐（Developer 默认 threads=2 片上直连 / 回退 workspace+vid） |
冲突处理：当 design.md 与参考实现冲突时：
- 优先参考实现：参考实现已验证通过，可信度高
- 记录差异：在代码注释中说明为何偏离 design.md
- 询问用户：重大差异需确认
步骤 3：生成实现代码
⚠️ 生成代码时必须遵守 references/ascend-constraints.md §2「Host 侧 Buffer 操作约束」：算子的核心计算逻辑全部在 kernel 内实现。host 侧对 NPU 张量只能做「只改元数据」的视图操作（
reshape/view/transpose/permute/expand）以及数据准备 / kernel 调用 / 结果验证；禁止改数据指针、禁止.contiguous()等真实重排、禁止改写 buffer 内容、禁止用新 buffer 作弊。拿不准时，一律放入 kernel。
⚠️ dtype 特化检查（支持多 dtype 的算子必须执行）：生成代码时必须检查每个支持的 dtype 是否走硬件加速路径。如果 dtype 回退标量路径，必须比较同宽 reinterpret、kernel 内 cast、record-aware DMA、块 DMA + UB-local 标量重排等候选。禁止把大张量降级成逐元素 strided GM load/store，也不能把“逐行 T.copy”误当成可完成任意转置。 UB 内局部标量 lowering 可以作为经最大 case 验证的 fallback；不得在 host 侧用
.to(dtype)/.contiguous()/torch.stack绕过。详见 references/coding-conventions.md §7。
⚠️ stride/shape 参数传递（kernel 需要地址偏移时）：优先作为
@tilelang.jit函数的 Python 参数传入（JIT 编译期常量），kernel 内用T.alloc_var累加偏移。不要打包成 int32 GM tensor 在运行时传入——会增加一次 GM→UB 搬运。详见 tilelang-api-best-practices/references/api-compute.md §4.10 Stride 参数作为 JIT 编译期常量。
⚠️ 数据重排实现门禁：GM↔UB 应使用尽可能大的块/二维
T.copy。标量循环只能用于 UB-local reorder，不能在大张量主路径中写成ub[i] = gm[base + i * stride]。 对连续 suffix record，必须按通用结构谓词聚合多条 record，而不是每条 record 发两次短 DMA。 生成后用get_kernel_source()检查每条 dtype/路径；若GetValue/SetValue对 GM 按 numel 展开，或 DMA transaction 估算达到数十万/百万级，必须重新设计。不要只执行上述门禁。凡是涉及数据布局变化，必须按 coding-conventions.md §6.1 的六步配方生成候选实现：识别连续 record → 按成本选路径 → 聚合搬运 → UB-local reorder → 必要时分阶段 → 数字验收。
完成候选后必须写 [REORDER-COST-AUDIT]，包含 GM pass、DMA transaction/平均字节、
GM 标量访问、逐元素 div/mod、active cores、每核串行任务和最大 case timeout。任一
指标无法估算时先读取 §6.1 补齐；大张量路径出现 numel 级 GM 标量访问或海量短 DMA
时不得进入精度验收。
基于 design.md 的 API 映射 + 参考示例的代码风格，生成两个文件：{op}.py（纯 kernel）与 test_{op}.py（golden + L0 + main，L1/L2/Boundary 留桩，从 {op}.py import kernel）。完整文件结构骨架与融合算子注意事项见 examples/code-skeleton.md。
写代码时遇到具体编码规范问题（Buffer 分配 / 索引一致性 / 同步 / 广播 / 测试模板）查 references/coding-conventions.md。
V 核并行化（按行切分、中间 buffer 索引一致性、CV 融合 V 核切分）查 references/vector-parallelism.md。
含 GEMM 或 CV 融合时查 references/gemm-cv-fusion.md（gemm_v0 初始化、NPU 分形限制、CV 融合必开的 4 个 pass_configs）。
Show full SKILL.md (340 more words)Show less
步骤 4：运行验证
本 skill 负责 L0 精度收敛，同时负责实现的最低性能可行性。先跑 L0：
python examples/{op}/test_{op}.py --level l0随后必须运行 design.md 中的性能可行性哨兵（即使它被标为 large/L1），为每个 case
设置明确 timeout。用户明确给出的失败或超时 case 必须全部实际运行。任一哨兵超时，
不得宣称生成完成：应修复搬运路径；若现有 API 无法满足，则返回 [DESIGN_ERROR]
并附 GM/DMA 成本证据。
测试数据准备同样属于 aclnn 审计范围：随机数、特殊值注入、dtype 转换和 golden
物理重排全部在 CPU 完成，然后只做一次 H2D；验证时只做 D2H。不得在 NPU 上调用
torch.rand/randint、in-place random、.contiguous() 或 golden 计算。报错中若出现
aclnnInplaceRandom，说明失败发生在测试输入准备，不是 kernel 内存不足或精度问题。
若测试规格包含 NaN/Inf，生成的测试必须采用位置敏感验证：
- 在 CPU 上用固定 seed 生成有限基础值和稀疏特殊值 mask，并保证至少一个特殊值、
一个有限值；
[nan, nan]不得直接退化为全 NaN。 - 数值容差判断前，分别要求 actual/golden 的 NaN、正 Inf、负 Inf mask 完全相等。
- mask 一致后，只在双方有限的位置计算 atol/rtol、matched ratio 和 max absolute error。
- 全 NaN/全 Inf 只能作为补充用例，不能作为唯一特殊值门禁；
torch.allclose(..., equal_nan=True)不能替代显式 mask 比较。
具体生成骨架见 references/coding-conventions.md §5。
L0 通过后，由
tilelang-op-test-design（场景 B）填充 L1/L2/Boundary 桩体，再--level all跑全量。 main 分发器与--level接口由本 skill 生成并保持稳定（模板见 code-skeleton.md），扩展时不改动。
如果报错，查阅 references/troubleshooting.md 进行排查：
| 错误类型 | 排查方向 | 详细参考 |
|---|---|---|
| 编译错误 | buffer 大小、API 参数、对齐 | troubleshooting.md §编译时错误 |
| 运行错误 | 索引越界、同步缺失 | troubleshooting.md §运行时错误 |
| 精度错误 | Golden 实现、输出形状 | troubleshooting.md §精度问题 |
遇到具体错误信息时，先查 references/troubleshooting.md ——本 skill 配套的疑难解答手册，覆盖编译错误（UB 内存不足 / threads / 动态循环边界）、运行错误（index OOB / valid_shape）、精度错误（dtype / atol 阈值）等常见场景的具体解决方案。
步骤 5：上库前检查清单
运行通过后，必须按 references/checklist.md 逐项检查。
⚠️ 首要检查：算子主要操作是否全部在 kernel 内实现（违反则立即修改，不得继续）
逐项检查前，先回顾生成的代码，按 references/ascend-constraints.md §2「Host 侧 Buffer 操作约束」的五条禁令逐条核对 host 侧代码（含 kernel 调用后的输出后处理路径）：
- 是否把输入/输出 tensor 重新绑定到别的 tensor（改了
data_ptr）后传入 kernel？ - 是否存在真实数据拷贝/重排？对 reshape 应验证 storage/stride 兼容性，不能把
is_contiguous() == False直接等同于一定复制；permute/transpose通常只改 metadata。 - 是否在 host 侧直接改写了 tensor 数据（
x[:] =、in-place_()、out=等）？ - 是否用「新建 buffer → host 侧处理 → 替换原 tensor」的方式作弊？
- 是否在 host 侧隐式触发了 aclnn 调用？重点检查：
torch.nn.functional.pad/cat/interpolate、torch.cat/stack、.to(dtype)dtype 转换、.clone()、以及输出侧切片+reshape（如y = y[:,:,:,:S]; y.reshape(shape)——切片后非 contiguous，reshape 隐式.contiguous()→aclnnCopy）。若需要从 padded 输出裁剪有效部分，应改为让 kernel 直接输出到与原始 shape 一致的 buffer（通过T.copy+pad_value处理尾块），host 侧无需切片。 - 是否对所有支持的 dtype 做了特化检查？用
get_kernel_source()确认每个 dtype 是否走硬件加速路径。标量回退的 dtype 是否在 kernel 内处理（cast 或逐行搬运），而不是在 host 侧用.to()/.contiguous()/torch.stack绕过？
任何一条命中，必须立即修改——把这些操作移入 kernel 内部，直到满足要求后才能继续后续检查。允许的 host 侧操作仅限：reshape/view/transpose/permute/expand 等只改元数据的视图操作，以及数据准备、kernel 调用、结果验证。
最容易踩坑的 4 项重点提醒：
| 关键项 | 说明 | checklist 编号 |
|---|---|---|
| Golden 实现一致 | 迁移算子必须使用原算子的 golden 实现 | #9 |
| tilelang.disable_cache() | 放在 __main__ 下方或 main() 内部 | #11 |
| 分层标记 + --level | L0/L1 打 [PRECISION_PASS/FAIL]、L2/Boundary 打 [BOUNDARY_PASS/WARN]；main 支持 --level；L0/L1 全过才 "Test Passed!"+exit 0 | #14-17 |
| 代码格式 | ruff check + ruff format --check 通过 | #18 |
4. Skill 反馈采集
算子开发流程跑完后触发，把"哪些 skill 没讲清楚 / 被现实打脸 / 凭经验补的内容"写到 .agents/skill-journal/。
⚠️ 触发权归属取决于调用模式（orchestrator 编排时不主动触发，单独调用时手动触发）。完整触发规则、枚举 skill、反思四问、写 journal schema、自检、完成报告见 references/skill-feedback.md。
子目录索引
- references/ascend-constraints.md — 算子 kernel 划分原则 + Host 侧 Buffer 操作约束（生成代码前/时必须遵守的强制规则）
- references/coding-conventions.md — Buffer 分配 / 索引 / 同步 / 广播 / 测试模板（写代码遇到具体规范时查）
- references/vector-parallelism.md — V 核并行化（用到 vid 切分时查）
- references/gemm-cv-fusion.md — GEMM 与 CV 融合 pass_configs（含 GEMM 或融合算子时查）
- references/checklist.md — 上库前检查清单（生成代码后逐项过）
- references/troubleshooting.md — 编译 / 运行 / 精度错误排查手册（遇到具体错误时查）
- references/skill-feedback.md — Skill 反馈采集流程（流程结束时查，orchestrator 模式跳过）
- examples/code-skeleton.md —
{op}.py（kernel）+test_{op}.py（测试）文件结构骨架
© tile-ai, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file
Files
SKILL.md and 8 other files (references) in .agents/skills/tilelang-op-develop of tile-ai/tilelang-ascend.
- SKILL.md
- examples/code-skeleton.md
- references/ascend-constraints.md
- references/checklist.md
- references/coding-conventions.md
- references/gemm-cv-fusion.md
- references/skill-feedback.md
- references/troubleshooting.md
- references/vector-parallelism.md
Open the folder on GitHubat commit 3d17c28
Compare with similar skills
Tilelang Op Develop next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.
| Skill | Stars | Used in | Tokens | Auto-check | Licence | Repo updated |
|---|---|---|---|---|---|---|
| Tilelang Op Develop this skilltile-ai/tilelang-ascend | 400 | — | ~2.6k | Automated safety check: Pass | MIT | |
| Figma Design System Builderwarpdotdev/warp | 65k | 2 repos | ~4.4k | Automated safety check: Pass | AGPL-3.0 | |
| Figma use_figma Plugin API Ruleswarpdotdev/warp | 65k | 4 repos | ~4.4k | Automated safety check: Pass | AGPL-3.0 | |
| MCP Developmentcoollabsio/coolify | 63k | 1 repos | ~949 | Automated safety check: Pass | MIT | |
| Design SystemOhh-889/skyroc | 795 | 11 repos | ~1.7k | Automated safety check: Pass | MIT | |
| Design Dnazanwei/design-dna | 1.9k | 1 repos | ~2.1k | Automated safety check: Pass | MIT |
Similar skills
Figma use_figma Plugin API Rules
warpdotdev/warp
Required groundwork before any use_figma call: the rules and reference files for running JavaScript in a Figma file through the Plugin API without common failures.
MCP Development
coollabsio/coolify
A skill your agent uses for Laravel MCP development. An agent skill from coollabsio/coolify.
Design System
Ohh-889/skyroc
Token architecture, component specifications, and slide generation.
Design Dna
zanwei/design-dna
Extract, define, and apply design DNA across three dimensions: design system (tokens), design style (qualitative feel), and visual effects (Canvas, WebGL, 3D, particles, shaders, scroll effects…
Figma Design to Code
warpdotdev/warp
Turns a Figma frame or component into production code that matches the design, using the Figma MCP server and the project's own design system.
More from tile-ai/tilelang-ascend
All 20 skills in this repoTilelang Submodule Pull
tile-ai/tilelang-ascend
Automatically pull tilelang repository and its third-party code.
Tilelang Debug Helper
tile-ai/tilelang-ascend
How to add debugging capabilities to TileLang Ascend example operators.
Tilelang A5 Sim Convert
tile-ai/tilelang-ascend
将 tilelang example 脚本转换为可在 A5 camodel 仿真器上直接运行的版本。输入脚本路径，输出一个新的 sim.py 文件，不覆盖原始文件。触发：仿真运行、camodel、A5 仿真、sim 模式、转换脚本为仿真、不需要 NPU 跑 kernel、simulate A5。
Tilelang Env Check
tile-ai/tilelang-ascend
TileLang-Ascend 环境检查与配置验证技能。检查代码仓库完整性、编译安装状态、环境变量配置，并运行简单测试验证环境。发现问题会自动调用相关 skill 进行修复，并按依赖顺序重新执行后续步骤。触发关键词："环境检查"、"检查环境"、"验证环境"、"环境配置"、"环境搭建"、"env check"、"check environment"、"verify…
Tilelang Op Test Design
tile-ai/tilelang-ascend
TileLang-Ascend 算子测试设计技能。支持多种场景：(1) 从 design.md 设计测试配置 (2) 从 examples/{op}/.py 补充测试 (3) 手动提供算子信息生成测试 (4) 测试覆盖率分析。理解算子实现逻辑后智能判断测试策略。触发：设计算子测试、生成测试用例、补充测试、测试覆盖率不足。
Tilelang Review
tile-ai/tilelang-ascend
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…
Related
Categories
Questions about Tilelang Op Develop
What does Tilelang Op Develop do?
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 Op Develop is an agent skill from tile-ai/tilelang-ascend.
When should I use Tilelang Op Develop?
Tilelang Op Develop fits situations like: tasks that involve Design tokens.
How do I install Tilelang Op Develop in Claude Code?
Run `npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a claude-code`. Or copy the skill folder (.agents/skills/tilelang-op-develop in tile-ai/tilelang-ascend) into .claude/skills/tilelang-op-develop in your project. Claude Code loads it when a task matches its description.
How do I install Tilelang Op Develop in Codex?
Run `npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a codex`. Or copy the skill folder (.agents/skills/tilelang-op-develop in tile-ai/tilelang-ascend) into .agents/skills/tilelang-op-develop in your project. Codex loads it when a task matches its description.
Can I use Tilelang Op Develop in Cursor, Gemini CLI or GitHub Copilot?
Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add tile-ai/tilelang-ascend --skill tilelang-op-develop -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/tilelang-op-develop, .gemini/skills/tilelang-op-develop, .github/skills/tilelang-op-develop and .opencode/skills/tilelang-op-develop in your project.
What does Tilelang Op Develop need to run?
Going by SKILL.md and its folder, Tilelang Op Develop needs the command-line tools its instructions call (ruff and python). Our summary lists: Python 3.
Does Tilelang Op Develop access the network?
SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.
Is Tilelang Op Develop safe to install?
Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.
What licence does Tilelang Op Develop use?
Tilelang Op Develop is published under the MIT licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.
How many tokens does Tilelang Op Develop use?
About 2.6k tokens (SKILL.md is roughly 11k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 13k tokens, read only when the agent opens those files.
What are the alternatives to Tilelang Op Develop?
Skills that share tags, products or a category with Tilelang Op Develop: Figma Design System Builder (warpdotdev/warp, 65k stars), Figma use_figma Plugin API Rules (warpdotdev/warp, 65k stars), MCP Development (coollabsio/coolify, 63k stars) and Design System (Ohh-889/skyroc, 795 stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.
Who maintains Tilelang Op Develop?
tile-ai (a GitHub organization) maintains it in tile-ai/tilelang-ascend, which has 400 GitHub stars. The repository holds 20 skills in this directory. The repository was last updated on October 8, 2026.
Source: tile-ai/tilelang-ascend on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.