ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
Creates GitHub pull requests with properly formatted titles that pass the check-pr-title CI validation. Use when creating PRs, submitting changes for review, or when the user says /pr or asks to create a pull request.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
Create new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.
ascend-ai-coding/awesome-ascend-skills
Agent Codex
Analyze blind comparison results to understand WHY the winner won and generate improvement suggestions.
ascend-ai-coding/awesome-ascend-skills
Agent Codex
Compare two outputs WITHOUT knowing which skill produced them.
ascend-ai-coding/awesome-ascend-skills
Agent Codex
Evaluate expectations against an execution transcript and outputs.
ascend-ai-coding/awesome-ascend-skills
Plugin Claude Code
A comprehensive knowledge base for Huawei Ascend NPU development, structured as distributed AI Agent Skills. Includes npu-smi device management, HCCL performance testing, and ATC model conversion.
ascend-ai-coding/awesome-ascend-skills
Instructions file CodexOpenCode
AGENTS.md instructions for ascend-ai-coding/awesome-ascend-skills, covering agents.md - awesome ascend skills agent guide, 1. 先读什么, 2. 仓库结构, 3. skill 类型 and 4. 如何使用本仓.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A guide to making custom PyTorch operations work in npugraphex graph mode on Huawei Ascend NPUs. It explains how to register Python custom operations and write the metadata needed to describe their outputs.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A troubleshooting guide for compile-time errors when PyTorch code is captured for Huawei Ascend NPUs using npugraphex. It covers failures in tracing, graph conversion, automatic differentiation, shape or metadata checking, and graph capture.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A first-step triage guide for npugraphex issues in PyTorch on Huawei Ascend NPUs. It collects the needed logs and classifies the problem as compilation, runtime, accuracy, or performance related.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
Skill "external-cannbot-graph-torch-npugraph-ex-knowledge" from ascend-ai-coding/awesome-ascend-skills, covering npugraphex 模式（aclgraph）, 模式概述, 核心规则, 文档获取方式 and torchair 主知识源.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A troubleshooting guide for PyTorch models running on Huawei Ascend NPUs with npugraphex, a graph-execution mode. It covers failures after the graph has been captured, such as operator, communication, memory, and device errors.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A coding template for creating npugraphex-mode examples from scratch. It prescribes a three-part layout: imports, one inference function, and the program entry point, with fixed variable names.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
An inference-optimization guide for replacing repeated PyTorch computation patterns with torchnpu fused operators on Huawei Ascend NPUs. A fused operator combines several calculations into one specialized operation.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A model-adaptation guide for running PyTorch inference in graph mode on Huawei Ascend NPUs. Graph mode compiles model operations into an execution graph, while a graph break leaves part of the model outside that graph.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
An optimization guide for KVCache in large language model text generation on Huawei Ascend NPUs. KVCache stores previously computed attention data so later tokens do not recalculate it.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A baseline guide for adapting and deploying PyTorch models for inference on Huawei Ascend NPUs. It takes a model from a Hugging Face link or local code and prepares it for the ModelRunner inference framework.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
An inference-optimization guide for using multiple execution streams in a PyTorch model on Huawei Ascend NPUs. Streams are ordered work queues that can sometimes run independent operations at the same time.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A guide for choosing how a model's work is split across Huawei Ascend NPU cards. TP splits tensor calculations, EP splits expert modules, and DP splits data batches across devices.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A PyTorch code-change workflow for splitting model inference across Ascend NPU devices according to an approved parallel configuration. It updates layers, communication groups, embeddings, output heads, YAML settings, and model weights.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A diagnostic guide for model-inference accuracy problems on Ascend NPUs when using PyTorch. It focuses on KVCache and FlashAttention issues, including Prefill and Decode behavior.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A PyTorch inference optimization workflow that adds torchnpu.npuprefetch to prefetch model weights on Ascend NPUs. Weight prefetch loads needed data ahead of its use.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A troubleshooting workflow for runtime errors when PyTorch models run inference on Ascend NPUs. It covers failures during model loading, setup, execution, and distributed communication.
ascend-ai-coding/awesome-ascend-skills
Skill Claude CodeCodex
A PyTorch inference adaptation workflow for enabling SuperKernel binary operator fusion on Ascend NPUs. Operator fusion combines work to reduce task-scheduling overhead, and this workflow applies only to graph mode, Atlas A3 hardware, and the decode stage.