Guides the design and documentation of custom Triton operators.
Install
mkdir -p .claude/skills/triton-operator-design && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/17155" && unzip -o skill.zip -d .claude/skills/triton-operator-design && rm skill.zipInstalls to .claude/skills/triton-operator-design
Activation
This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.
生成适用于 Ascend NPU 的 Triton 算子需求文档。当用户需要设计新的 Triton 算子、编写算子需求文档、进行算子性能优化设计时使用。核心产出：功能定义、API 接口、Tiling 策略、Kernel 实现方案。Key capabilities
- →Analyze requirements for Triton operators, including functional definitions and mathematical formulas
- →Design API interfaces for new Triton operators
- →Define input/output constraints and hardware limitations for operators
- →Develop Tiling strategies for optimal performance on Ascend NPU
- →Propose Kernel implementation schemes for Triton operators
- →Ensure adherence to specific memory alignment and precision requirements
How it works
The skill guides through a workflow from requirements analysis to feature implementation, loading specific reference resources at each stage to generate a Triton operator design document.
Inputs & outputs
When to use triton-operator-design
- →Design custom NPU operator
- →Optimize kernel tiling
- →Document operator specs
About this skill
Triton 算子需求文档生成
工作流
- 需求分析 → 产出：功能定义、数学公式、竞品对比
- 原型设计 → 产出：API 接口定义
- 规格约束 → 产出：输入输出约束、硬件限制
- 特性实现 → 产出：Tiling 策略、Kernel 实现方案
参考资源加载
| 阶段 | 必须加载 | 不要加载 |
|---|---|---|
| 需求分析 | ascend-terminology.md | template, tiling-strategies |
| 原型设计 | triton-api-reference.md | tiling-strategies |
| 特性实现 | tiling-strategies.md, triton-operator-template.md | ascend-terminology |
MANDATORY：阶段 4 前，完整阅读 tiling-strategies.md 和 triton-operator-template.md，不设行数限制。按模板格式输出文档。
关键术语
- GM：全局内存（DDR），UB：Vector Core 高速缓存（192KB），L1：Cube Core 缓存（~1MB）
- AI Core：A2/A3 有 24 个，含 1 Cube + 2 Vector
- 归约操作必须升精度到 FP32
绝对不要做的事
- ❌ 使用模糊术语（"适当切分"、"合理分配"）— 必须给出具体计算方法
- ❌ 忽略 UB 大小（192KB）和对齐要求（32B）
- ❌ 不区分 Vector Core（向量计算）和 Cube Core（矩阵计算）
- ❌ 不标注数据流图中的数据类型和 GM↔UB 传输
- ❌ 归约操作不说明升精度策略
常见陷阱
| 陷阱 | 症状 | 解决 |
|---|---|---|
| UB 超限 | 方案不可实现 | 计算缓冲区总大小 < 192KB |
| 内存未对齐 | 硬件报错 | UB 缓冲区 32B 对齐，单值缓冲区分配 32B |
| 精度损失 | FP16 结果不准 | 归约前升 FP32，完成后降精度 |
| Tiling 不合理 | 性能差/大 shape 不支持 | 按维度切分，避免跨 Core 数据依赖 |
When not to use it
- →When designing operators for platforms other than Ascend NPU
- →When the task does not involve Triton operators
- →When the request is for general software design not specific to NPU operators
Limitations
- →Requires specific reference resources to be loaded at different stages
- →Strict adherence to template format for output document
- →Prohibits vague terminology, requiring concrete calculation methods
How it compares
This workflow provides a structured, NPU-specific approach to Triton operator design, ensuring consideration of hardware constraints and optimization strategies from the outset, unlike general software development processes.
Compared to similar skills
triton-operator-design side by side with the closest alternatives in the catalog.
| Skill | Installs | Updated | Safety | Difficulty |
|---|---|---|---|---|
| triton-operator-design (this skill) | 0 | 3mo | No flags | Advanced |
| langchain-architecture | 8 | 3mo | Review | Intermediate |
| ai-agents-architect | 5 | 6mo | No flags | Advanced |
| gguf-quantization | 6 | 7mo | Review | Intermediate |
Try saying
Example prompts that trigger this skill in your AI assistant.
More by Ascend
View all by Ascend →You might also like
langchain-architecture
wshobson
Design LLM applications using the LangChain framework with agents, memory, and tool integration patterns. Use when building LangChain applications, implementing AI agents, or creating complex LLM workflows.
ai-agents-architect
davila7
Expert in designing and building autonomous AI agents. Masters tool use, memory systems, planning strategies, and multi-agent orchestration. Use when: build agent, AI agent, autonomous agent, tool use, function calling.
gguf-quantization
davila7
GGUF format and llama.cpp quantization for efficient CPU/GPU inference. Use when deploying models on consumer hardware, Apple Silicon, or when needing flexible quantization from 2-8 bit without GPU requirements.
pytorch-fsdp2
Orchestra-Research
Adds PyTorch FSDP2 (fully_shard) to training scripts with correct init, sharding, mixed precision/offload config, and distributed checkpointing. Use when models exceed single-GPU memory or when you need DTensor-based sharding with DeviceMesh.
huggingface-accelerate
davila7
Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for DeepSpeed/FSDP/Megatron/DDP. Automatic device placement, mixed precision (FP16/BF16/FP8). Interactive config, single launch command. HuggingFace ecosystem standard.
llm-app-patterns
davila7
Production-ready patterns for building LLM applications. Covers RAG pipelines, agent architectures, prompt IDEs, and LLMOps monitoring. Use when designing AI applications, implementing RAG, building agents, or setting up LLM observability.