awesome-ascend-skills

AI for Science skill set for Huawei Ascend NPU. Includes a main routing skill for profiling collection, model migration, and TensorFlow route selection, plus specialized leaf skills for Boltz2, BoltzGen, DeepFRI, DiffSBDD, GENERator, OligoFormer, ProteinBERT, TensorFlow Community deployment, TF-to-PyTorch conversion, and torch_npu profiling workflows.

191 skills

官方推荐的 AI for Science 安装包。适用于 AI for Science 场景下的 Profiling 采集、模型迁移、TensorFlow/Keras 路线选择与相关子方向能力。

191 skills

官方推荐的昇腾基础安装包。适合新同学和通用项目起步，包含设备检查、容器环境与 PyTorch NPU 基础能力。建议作为所有场景的第一步安装入口。

191 skills

官方推荐的昇腾推理安装包。适用于模型转换、量化、推理部署与评测场景，包含 ATC、vLLM-Ascend、ais-bench、msmodelslim、Diffusers 与 Wan 适配相关 skills。

191 skills

官方推荐的昇腾算子开发与迁移安装包。适用于 AscendC 算子开发、torch_npu 自定义算子接入、Triton-Ascend 迁移与算子基准测试场景。

191 skills

官方推荐的昇腾 Profiling 与性能分析安装包。适用于 Profiling 数据采集、瓶颈分析、训练性能诊断与 MFU 分析场景。

191 skills

官方推荐的昇腾训练安装包。适用于分布式训练、通信测试、MindSpeed-LLM、MindSpeed-MM 与 VERL 训练链路，包含 HCCL、torch 通信测试、VERL quickstart 与 VERL msprobe 精度采集相关 skills。

191 skills

Diffusers 昇腾 NPU 技能集，包括环境配置、权重准备、Pipeline 推理等。当用户需要在昇腾 NPU 上运行 Diffusers 推理时使用。

191 skills

AISS-TilingSolver：Ascend C 算子 Tiling 参数自动求解工具（MatMul / Vector）。下载安装、构造 JSON 输入、运行求解、结果解读与故障排查。

191 skills

NPU 模型推理优化原子 Skills（11 个）。覆盖框架适配、并行策略、KVCache/FA、融合算子、图模式、精度/运行时调试、多流、预取和 SuperKernel。

191 skills

Ascend C 代码审查 Skills（2 个）。代码审查规范和 API 文档检索。

191 skills

Ascend C Kernel 直调算子开发 Skills（40+ 个）。API 检索、Tiling 设计、精度调试、性能采集、环境检查、代码模板等。

191 skills

PyPTO 算子开发 Skills（8 个）。API 探索、Golden 生成、需求理解、算子设计与开发、性能调优、精度调试。

191 skills

TileLang 算子开发 Skills（5 个）。API 最佳实践、编程模式指南、算子设计、算子开发、代码审查。

191 skills

PyTorch 昇腾 NPU npugraph_ex Skills（6 个）。覆盖图模式知识、代码模板、问题分诊、编译错误、运行时错误和自定义算子入图。

191 skills

Triton-Ascend 算子代码生成与优化 Skills（6 个）。任务提取、算法设计、代码生成、功能验证、性能优化，以及 NPU 架构参考。

191 skills

从 gitcode-ascend 同步的 Ascend 技能集，包含 91 个技能： - analyse-coverage: 分析测试覆盖率盲区，生成覆盖率分析报告 - arxiv-recommendation-npu: 自动化推荐系统论文发现流水线。抓取 arxiv 推荐论文，检测源码，生成待迁移任务清单，由 npu-model-migration skill 完成 NPU 适配。 - ascend-inference-repos-copilot: 昇腾（Ascend）推理生态开源代码仓库智能问答专家旨在为 vLLM、vLLM-Ascend、MindIE-LLM、MindIE-SD、MindIE-Motor、MindIE-Turbo 以及 msM - ... 等 88 个技能

191 skills

从 mindstudio 同步的 Ascend 技能集，包含 9 个技能： - ascend-profiler-db-explorer: 面向 Ascend PyTorch Profiler / msprof DB（如 ascend_pytorch_profiler*.db、msprof_*.db）的 SQL 分析技能。将自然语言问题（ - cluster-fast-slow-rank-detector: 专门用于 Ascend 集群 Profiling 性能数据的“快慢卡”诊断专家技能。当用户提供【集群性能数据目录/路径】并要求分析【快慢卡】、【慢节点】、【负载不均衡】或【集群瓶颈】时，必须触发此技能 - document-ux-review: 当用户希望你像第一次接触项目的人一样，真实按仓库的 README、安装文档或 quick start 跑一遍，并判断“新人能不能走通”“文档是否可用”“哪里会卡住”“安装/启动说明是否对新手友好”时， - ... 等 6 个技能

191 skills

基于华为昇腾社区论坛开发者反馈内容，按时间范围抓取帖子并识别特定领域的易用性、故障和报错类问题，用于产品改进分析。

191 skills

MindSpeed-LLM skill set for Huawei Ascend NPU, covering the full training pipeline: environment setup, data preprocessing, weight conversion, distributed training, and end-to-end orchestration. Use when deploying MindSpeed-LLM training on Ascend NPU.

191 skills

MindSpeed-MM 多模态大模型训练技能集，用于华为昇腾 NPU。覆盖环境搭建、权重转换、多模态理解模型训练、多模态生成模型训练和端到端流水线。

191 skills

Huawei Ascend NPU server model quantization tool skills, using msmodelslim for quantization compression, the tool is specifically designed for LLM quantization acceleration on Ascend NPU cards. It supports INT4/INT8/INT16 quantization and provides a complete quantization workflow, including environment checking, dynamic scheme query, container deployment, quantization execution, etc., and supports quantization tuning, adapter adaptation, etc. Interactive experience with controllable processes.

191 skills

Profiling skill set for Huawei Ascend NPU performance analysis. Analyzes step_trace_time.csv files to identify hostbound, computing, and communication bottlenecks. Use when performing performance analysis and optimization on Ascend NPU.

191 skills

Guides migration of traditional models from CPU/GPU to Ascend NPU. Invoke when user needs to adapt PyTorch code for Ascend devices, set up NPU environment, or verify migration correctness.

4 skills

AI model evaluation tool for Ascend NPU. Supports accuracy evaluation (text, multimodal datasets), performance evaluation (latency, throughput, stress testing), vLLM/Triton inference, 15+ benchmarks (MMLU, GSM8K, MMMU, BFCL), multi-turn dialogue, Function Call, and custom datasets.

1 skill

昇腾虚拟化实例（AVI）和vNPU管理技能。用于查询/设置AVI模式、管理vNPU实例（创建/销毁）、查询vNPU配置恢复状态等操作。支持安全确认机制，在执行修改操作前检查设备状态。

1 skill

End-to-end AscendC custom operator development for Ascend NPU in an ascend-kernel (csrc/ops + build.sh + torch_npu PyTorch custom op) project. Drives a new operator from a name and a math spec through project init, two-level tiling design, op_host/op_kernel code generation and framework registration, compile/install/debug, PyTorch-style API docs, precision evaluation and root-cause debugging, torch_npu.profiler performance benchmarking, performance optimization, and security code review. Use to design, generate, build, test, document, and tune an AscendC operator.

1 skill

当用户需要对华为昇腾 NPU 进行硬件层面的管理、测试或诊断时使用，包括状态/温度/利用率查询、h2d/d2h/d2d/p2p 带宽、TFLOPS/TOPS/功耗基准、硬件故障诊断、压力测试与异常 NPU 卡复位；不要用于写 NPU 代码、安装框架、训练调试或 Docker 配置。

1 skill

Create Docker containers for Huawei Ascend NPU development with proper device mappings (davinci_manager, devmm_svm, hisi_hdc) and volume mounts (driver, sbin, home). Use when setting up Ascend development environments in Docker, running CANN applications in containers, or creating isolated NPU development workspaces.

1 skill

通用 PyTorch 项目 Ascend NPU 迁移可行性分析。系统化扫描代码库中的 CUDA/GPU 依赖，按 7 大域分类评估（设备层、注意力机制、自定义算子、分布式通信、精度策略、第三方依赖、编译加速），并基于 Wan2.2 实际迁移经验提供逐项替代方案。适用于评估任何 DiT/Transformer 类模型（视频生成、LLM、多模态等）在昇腾 NPU 上的运行可行性与迁移工作量估算。

1 skill

用于指导 Ascend C（CANN / xpu_kernel）算子开发完成后接入 PyTorch：与 torch_npu 对接、Host 侧注册与构建、Python 暴露与冒烟测试。涵盖算子工程内轻量化接入、必要时参考 op-plugin、Pybind 快速联调、复用图算子封装等。适用于 Ascend C、算子接入、torch_npu、NPU 自定义算子、op-plugin、xpu_kernel、cpp_extension、自定义算子等场景。

1 skill

Complete toolkit for Huawei Ascend NPU model conversion and end-to-end inference adaptation. Workflow 1 auto-discovers input shapes and parameters from user source code. Workflow 2 exports PyTorch models to ONNX. Workflow 3 converts ONNX to .om via ATC with multi-CANN version support. Workflow 4 adapts the user's full inference pipeline (preprocessing + model + postprocessing) to run end-to-end on NPU. Workflow 5 verifies precision between ONNX and OM outputs. Workflow 6 generates a reproducible README. Supports any standard PyTorch/ONNX model. Use when converting, testing, or deploying models on Ascend AI processors.

1 skill

HuggingFace Diffusers 环境配置指南，用于华为昇腾 NPU。覆盖 CANN 版本检测、PyTorch + torch_npu 安装、Diffusers 库安装及环境验证。当用户需要在昇腾 NPU 上配置 Diffusers 环境时使用。

1 skill

Automates GitCode open-source repo merge workflow: commit → push → issue → PR → pipeline → review → /lgtm & /approve → merge. Invoke when user wants to submit code to a GitCode open-source repository.

1 skill

GitHub Issue root cause analysis skill. Analyze specified GitHub issues to identify possible root causes by examining issue content, code repository, and related resources. Use when user asks to analyze issue root cause, investigate issue, RCA analysis, or troubleshoot GitHub issues. Provides investigation directions with probability estimates and explicitly states when root cause cannot be determined.

1 skill

Analyze closed GitHub issues to create troubleshooting case studies with root cause analysis and lessons learned. Use when summarizing resolved issue troubleshooting process, creating issue case studies, documenting problem-solving methodology, extracting lessons from closed issues, or building knowledge base from issue history.

1 skill

HCCL (Huawei Collective Communication Library) performance testing for Ascend NPU clusters. Use for testing distributed communication bandwidth, verifying HCCL functionality, and benchmarking collective operations like AllReduce, AllGather, AlltoAll.

1 skill

模型层 Tensor 打点与精度对比工具。用于在模型 forward 过程中捕获模型各层中间 tensor，实现 GPU/NPU 精度对比调试。支持 vLLM、SGLang 推理框架。When to use：When you need to debug precision issues between GPU and NPU, or validate layer-wise tensor outputs during inference.

1 skill

Automatically fetch InferenceX benchmark data and generate daily performance reports for LLM inference on various hardware (NVIDIA, AMD, etc.). Supports email delivery, data change detection, and 8k1k sequence length performance analysis. Use when needing to track LLM inference performance trends, compare hardware configurations, or monitor benchmark updates.

1 skill

指导并自动化完成昇腾 NPU 上 MindSpeed-LLM 训练的 Profiling 数据采集。支持配置并运行带 Profiling 的模型训练，包括 CPU 采集、内存采集、不同采集级别（level0/level1/level2）和自定义 step 范围。生成的 Profiling 数据可用 MindStudio Insight 进行性能分析。当用户需要在模型训练中采集 Profiling 数据、进行训练性能分析、或执行 性能数据采集/Profiling采集 时触发。触发关键词：profiling、性能分析、性能数据采集、Profiling采集、训练框架profiling、MindSpeed-LLM profiling。

1 skill

指导并自动化完成昇腾 NPU 上 MindSpeed-MM 模型训练的 Profiling 数据采集。通过识别训练脚本，自动找到配置文件并启用内置 Profiler，支持静态采集（指定起止 step）和动态采集（运行时动态开关），支持 Megatron 引擎（tools.json）和独立 FSDP2 引擎（YAML）两种配置方式。生成的 Profiling 数据可用 MindStudio Insight 进行性能分析。当用户需要在多模态模型训练中采集 Profiling 数据、进行训练性能分析、或执行 性能数据采集/Profiling采集 时触发。触发关键词：profiling、性能分析、性能数据采集、Profiling采集、多模态训练profiling、MindSpeed-MM profiling。

1 skill

Directly launch Docker containers on Ascend NPU servers for users. Collect requirements and execute docker run commands with proper NPU device mounting, networking, volume mounts, and environment variables. Use when users need to: (1) Start a Docker container on an NPU server, (2) Create containers with NPU device access, (3) Configure and launch containers with specific networking or mount requirements.

1 skill

Ascend NPU operator benchmark for repeat_interleave-style latency testing. Supports SSH remote, local Conda, and Docker environments with CANN toolkit detection and torch_npu runtime checks. Uses existing environments only and does not install or repair dependencies.

1 skill

Huawei Ascend NPU npu-smi command reference. Use for device queries (health, temperature, power, memory, processes, ECC), configuration (thresholds, modes, fan), firmware upgrades (MCU, bootloader, VRD), virtualization (vNPU), and certificate management.

1 skill

Migrate any HuggingFace model to Ascend NPU torchair graph mode (torch.compile) and benchmark it for accuracy and performance against NPU eager and CPU eager. Use when running, compiling, or benchmarking HF models (vision, text, image-text encoders such as SigLIP2, DINOv3, ViT, CLIP, Qwen-VL, SAM) on Ascend 910B/CANN with torch_npu and torchair; when a torch.compile graph-mode run on NPU fails (Dynamo TorchRuntimeError, unsupported op, interpolate/contiguous errors); or when comparing torchair vs npu_eager vs cpu with cosine similarity, max abs diff, and p50/p95/p99 latency.

1 skill

使用 torch_npu.profiler 在非 MindSpeed-LLM 与非 MindSpeed-MM 的训练/推理脚本中采集 Ascend NPU Profiling 数据。覆盖 level0/level1/level2 采集级别、训练循环/单次推理/指定代码段三种方式，支持 .py 和 .sh 脚本路径。当用户需要性能采集、性能分析、查看算子耗时、定位训练瓶颈时使用。触发关键词：profiling、性能采集、性能分析、算子耗时、瓶颈定位、torch_npu.profiler。

1 skill

Auto run NPU inference/training tests on a remote SSH server with Docker container(specified by user,default: vllm-ascend). Use when the user asks to test models on NPU, run inference on Ascend devices, or deploy models to an SSH server.

1 skill

Guide for connecting to remote servers via SSH, managing containers, and executing commands remotely. Covers SSH connection setup with multiple authentication methods (SSH key, sshpass, Paramiko, Fabric, tmux interactive), Docker container connection and command execution, file transfer, and troubleshooting. Use this skill whenever users need to: (1) Connect to a remote server or VPS via SSH, (2) Execute commands on remote machines, (3) Transfer files to/from remote servers, (4) Connect to Docker containers on remote hosts, (5) Set up SSH password authentication tools. Even if the user just says 'help me run something on my server' or 'connect to my machine', this skill applies.

1 skill

Automates verl msprobe precision dump workflows on Ascend/NPU: training profiler (global_profiler + precision_debugger stages), rollout inference dump (vLLM/SGLang), and train-inference consistency (engine patch + PROMPTS_ONLY). Use when users mention verl dump, msprobe, precision_debugger, rollout dump, or train-inference consistency.

1 skill

评估 CANN 算子仓「进阶教程 / 开发指南」类文档质量:通读文档 + 对照算子代码静态查证(默认不跑),找出 漏讲 / 对不上代码 / 过时 / 概念讲错,产出带证据与改进建议的体检报告(MD + HTML)。用于文档审稿、文档对照代码、tutorial 体检、文档可信度核查。

1 skill

华为昇腾 Ascend Extension for PyTorch (torch_npu) 的环境检查、部署与能力指引。在用户使用 @torch_npu、昇腾 NPU、CANN、或需要将 PyTorch 迁移到 NPU 时自动应用；当用户使用 @torch_npu_doc 时，基于 skill 的 reference 文档提供项目内中文文档能力说明。

1 skill

通过 PyTorch torch.distributed 接口测试昇腾 NPU 通信算子性能。支持指定任意 tensor shape、dtype，使用 torchrun 启动，贴近真实训练场景的通信算子测试与性能分析。Use for testing collective communication operators (AllReduce, AllGather, ReduceScatter, etc.) with specific tensor shapes via torch.distributed on Ascend NPU.

1 skill

大模型训练 MFU (Model FLOPs Utilization) 计算工具。支持 Dense 和 MoE 模型，GQA 和 SwiGLU FFN。包含 NVIDIA GPU 和 Ascend NPU 硬件峰值 FLOPS 参考，默认使用 Ascend910B2 (353 TFLOPS)。提供详细的 FLOPs 分析和性能报告。

1 skill

将 GPU/CUDA Triton 算子迁移为 Triton-Ascend，或将 Python/PyTorch 算子改写为可在 Ascend NPU 上运行的 Triton-Ascend 实现，并在发现明确优化空间时直接输出优化后的代码、最小验证脚本和排障说明。用户只要提到 昇腾、Ascend、NPU、triton-ascend、Triton 算子迁移、PyTorch 算子改写、coreDim、UB overflow、1D grid、物理核绑定、block_ptr、stride、访存对齐、mask 性能、dtype 退化、算子优化，或者直接问“这个 skill 怎么用”“怎么在命令行里跑”“怎么在容器里执行迁移/验证”，就应优先使用本 skill，即使用户没有明确说“写 skill”或“做迁移”。

1 skill

Generates an executable, end-to-end VERL reinforcement learning quickstart runbook for Ascend/NPU (docker image, dataset preprocessing, model setup, main_ppo training, and examples/run_*.sh flow). Use when users mention verl, quickstart, PPO/GRPO, gsm8k, run_*.sh, Ascend, or NPU training.

1 skill

vLLM Ascend plugin for LLM inference serving on Huawei Ascend NPU. Use for offline batch inference, API server deployment, quantization inference (with msmodelslim quantized models), tensor/pipeline parallelism for distributed serving, and OpenAI-compatible API endpoints. Supports Qwen, DeepSeek, GLM, LLaMA models with Ascend-optimized kernels.

1 skill

Deploy vLLM inference services on Ascend NPU servers with automatic model detection and optimized configuration. Supports local and remote deployment across bare metal, containers, and Docker images. Handles model discovery, quantization auto-detection, tensor parallelism configuration, graph/eager mode selection, and service health verification. Use when users need to: (1) Start or deploy vLLM server on NPU, (2) Launch LLM inference service, (3) Configure multi-card tensor parallel deployment, (4) Enable speculative decoding (Eagle) or quantization, (5) Run vllm offline batch inference, (6) Check or test vLLM service status.

1 skill

Interactive online benchmark orchestrator for vLLM inference services using `vllm bench serve`; supports single benchmarks, batch execution, result aggregation, and auto-optimization for concurrency/throughput under TTFT, TPOT, P99 and success-rate SLO constraints.

1 skill

Track daily PRs and Issues from vllm-project/vllm and vllm-project/vllm-ascend, filter by model (DeepSeek/Qwen/GLM/MiniMax/Kimi) and tech topics (PD disaggregation, MTP, quantization, graph mode, performance), analyze with LLM, and generate a Markdown report. Use when user wants vllm daily tracker, PR/Issue digest, or Ascend inference ecosystem monitoring.

1 skill

Comprehensive guide for adapting Wan series video generation models (Wan2.2/2.1/1.3) and similar diffusion-based video frameworks to Huawei Ascend NPU.

1 skill