tile-aiOfficial·11 Agent Skills Included

tilelang-ascend

Design, build, test, and tune Ascend NPU compute kernels

Generates design documents, kernel code, and layered test suites for high-performance operators on Huawei Ascend NPUs using the TileLang Python DSL. Covers compiler pass analysis, pass design and implementation, performance tuning, debugging, and environment verification. Removes guesswork around memory hierarchies, synchronization, and hardware constraints so kernels compile and pass precision checks faster.

346|151

↗

npx skills add tile-ai/tilelang-ascend --all -g -y

Available:

Routes operator development requests to the right sub-skills and enforces core rules like verifying APIs against examples, respecting the Ascend memory hierarchy, and never reporting unverified results as complete.

All Skills in This Repository (11)

Pure Emerald Level Indicators

📦 In Repo

tile-ai

tilelang-pass-analyzer

Analyze TileLang Pass functionality and generate structured Markdown reports.

tilelang-op-design

Generate TileLang-Ascend operator design documents from requirements.

tilelang-pass-workflow-analyzer

Analyzes TileLang-Ascend Pass pipeline relationships and execution order for placement recommendations.

tilelang-op-generate

Generate TileLang-Ascend operator code and tests from design.md specifications.

tilelang-review-skill

Validate Python and C++ code formatting against CI rules and generate JSON/Markdown reports.

tilelang-mode-guide

Guide TileLang Ascend mode selection and pass_config configuration for NPU kernels.

tilelang-github-operations

Install GitHub CLI and create pull requests via authenticated workflows.

tilelang-debug-helper

Configures GDB and VSCode debugging for mixed Python-C++ TileLang Ascend examples.

tilelang-error-fixer

Diagnose and repair TileLang-Ascend Pass crashes via GDB backtraces, IR dumps, and AOT tests.

tilelang-submodule-pull

Pull tilelang submodules and third-party code with retries and logging.

tilelang-api-best-practices

Provide best practices for writing Ascend NPU kernels with TileLang API.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install tilelang-ascend skills?▼

Run `npx skills add tile-ai/tilelang-ascend --all -g -y` in your terminal to install all skills in this suite globally.

How to write a custom Ascend NPU kernel?▼

Describe your operator's math formula and tensor shapes, and the skills generate a design document, kernel implementation, and precision tests based on verified examples in the repository.

How to optimize kernel performance on Ascend NPU?▼

The performance tuning skill profiles your kernel with msprof, identifies bottlenecks by operator type, and applies optimizations like double buffering and pipelining one step at a time.

Can it help debug TileLang compile errors?▼

Yes. Dedicated skills diagnose compile, runtime, and precision errors, support GDB and IR dump analysis, and suggest concrete fixes instead of blind rewrites.

Do I need deep NPU hardware knowledge to use it?▼

No. The skills encode Ascend-specific constraints like memory hierarchy rules and block size limits, so you can work from plain operator descriptions.