Awesome Ascend Skills
Over 200 skills for development on Ascend NPU chips: device diagnostics, model conversion, training and profiling.
- 182 ★ on the repository
- Last commit
Extended access
Why this level
- Diagnostic and launch skills such as npu-smi and the remote-server guide run commands on the device and over SSH on a remote server
- A sync script downloads and overwrites external skills from third-party repositories without the user controlling the edit's content
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/awesome-ascend-skillsDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- Diagnostic and launch skills such as npu-smi and the remote-server guide run commands on the device and over SSH on a remote server
- A sync script downloads and overwrites external skills from third-party repositories without the user controlling the edit's content
This is third-party code. Review the repository files before installing.
What it does
The repository collects skills for Huawei Ascend NPU development as SKILL.md files readable by Claude Code, OpenCode, Cursor, Trae and Codex. All skills live under skills/ by domain: base environment, inference, training, profiling, operator development, AI for Science and agent tooling, plus an external/ folder of skills synced from other Ascend repositories. Installing uses npx skills add with a domain's URL and the -s '*' flag; the base-domain install adds npu-smi device diagnostics, Docker configuration, a PyTorch-on-NPU check and a remote-connection guide. A single skill installs by name with npx skills add ascend-ai-coding/awesome-ascend-skills -s npu-smi, and without a choice the README gives a ready decision tree for which bundle to install first. The domain bundles cover the whole NPU workflow: ascend-inference for ONNX-to-.om conversion through ATC and running vLLM-Ascend, ascend-training for HCCL bandwidth tests and the full MindSpeed-LLM training pipeline, ascend-profiling for finding compute, communication and idle bottlenecks, and ascend-ops for AscendC operator development and porting Triton kernels. Skills also install by hand, by copying the chosen folder into a project's .agents/skills/ or a supported tool's global skills folder.
Who it is for. Engineers porting models and training to Huawei Ascend NPU hardware who want routine checks, conversion and profiling through ready skills instead of starting from scratch.
Good fit when
- A PyTorch project needs porting to Ascend NPU, and the first step is estimating effort and CUDA dependencies
- An ONNX model needs converting to .om format and comparing accuracy and performance against the original
- MindSpeed-LLM training has slowed down, and profiling is needed to find where time is lost: compute, communication or idle time
- HCCL needs testing for bandwidth and standard collective operations like AllReduce and AllGather before a training run
- A developer is writing a custom AscendC operator and needs a project template with tiling design and op_host and op_kernel code generation
Not a fit when
- There is no Ascend NPU hardware in the project: the skills target that architecture and do not transfer to CUDA or other NPUs without changes
- A single universal skill for every case is wanted: the README explicitly tells newcomers to install only ascend-base first, not the whole repository
- The agent tool cannot read the SKILL.md format: manual copying of the needed folder is the only option left
- The task does not fit any of the base, inference, training, profiling, ops or AI for Science domains the repository is organized around
Example request
Install the npu-smi skill from awesome-ascend-skills and check the health, temperature and load of every NPU on this machineLimitations
- The README and most skill descriptions are in Chinese; there is no English README in the repository.
- The external/ folder is synced automatically from third-party repositories such as cannbot, and the rules say not to edit those files by hand.
- Third-party rules are off by default until a machine administrator enables them explicitly, which also covers part of the external/ content.
- Installing with npx skills add pulls in a whole chosen domain or the '*' flag; there is no finer granularity between a single skill and a whole domain.
How to disable. In Claude Code run /plugin uninstall <domain>@awesome-ascend-skills, or delete the installed folder from .agents/skills/ by hand if it was installed without the marketplace.
FAQ
Where to start installing if Ascend NPU is new to you?
The README tells newcomers to install only ascend-base first, then add ascend-inference, ascend-training, ascend-profiling or ascend-ops for a specific task, not the whole repository at once.
How is ascend-profiling different from ascend-ops?
ascend-profiling collects and analyzes profiling data for an already-running model or training job, looking for compute, communication and idle bottlenecks. ascend-ops is for writing or porting the operator itself, for example with AscendC or a Triton port.
Can one skill be installed without the whole domain?
Yes, npx skills add ascend-ai-coding/awesome-ascend-skills -s npu-smi (or another name) installs exactly one skill without the rest of its domain.
What is in the external folder, and can it be edited?
These are skills synced automatically from third-party repositories, such as cannbot. The repository's rules explicitly ask not to edit these files by hand, since the next sync overwrites them.
Do these skills work with tools other than Claude Code?
Yes, the README explicitly lists OpenCode, Cursor, Trae and Codex as supported tools for reading the SKILL.md format, and the npx skills add install works the same way for all of them.
Discussion
No questions or reviews yet.
Related
Skills24 skills and a reference for an agent writing Triton-Ascend kernels for Ascend NPUs: tiling, autotuning, accuracy checks
Over 1,100 ready-made skills for Claude Code, Codex, Gemini CLI and Cursor, plus an MCP server that picks a skill for the task
skills-npm links agent skills shipped in your dependencies in node_modules and listed in the skills field of package.json
A set of independent skills against drift in large codebases: hard-cut, root-cause tracing, lean testing and CI right-sizing.
The official catalog of NVIDIA agent skills: CUDA, accelerated libraries, RAG, robotics and simulation through coding agents
Dimillian Skills
Skills
A set of 16 engineering skills for agents: Apple platform work, macOS packaging, React performance, code review and refactoring
README in short
The README opens with a decision tree for "what to install first" and a table of the official bundles, ascend-base, ascend-inference, ascend-training, ascend-profiling, ascend-ops and ascend-ai-for-science, with notes on who each one fits. It then lists every skill by domain with paths to its SKILL.md, a separate section on externally synced skills from cannbot, an explanation of how the skill build and compiler work, and governance rules for adding a new skill to the repository.