mindspore-aiOfficial·117 Agent Skills Included

akg

Automated GPU and NPU kernel generation, verification, and tuning

Generates, verifies, and optimizes compute kernels for CUDA GPUs, Ascend NPUs, and CPUs across Triton, CUDA C, TileLang, C++, and PyPTO. Eliminates manual kernel writing, tiling guesswork, and repetitive correctness testing through search-based and evolutionary optimization workflows. Includes DSL reference guides, hardware constraints, error-fix patterns, and benchmark harnesses to produce fast, correct operators with less effort.

258|48

↗

npx skills add mindspore-ai/akg --all -g -y

Available:

Tells the agent how to route kernel generation and optimization requests to the correct sub-skills, enforce valid backend/DSL/architecture parameter values, and follow environment setup and verification rules.

All Skills in This Repository (117)

Pure Emerald Level Indicators

📦 In Repo

mindspore-ai

task-constructor

Extract PyTorch/Triton operator implementations into single-file KernelBench tasks.

performance-summary

Summarize and compare operator performance results from result.json files.

kernel-workflow

Automate end-to-end AI kernel generation and optimization workflows across Triton, CUDA C, C++, and TileLang.

cuda-c-basics

Explain and implement basic CUDA-C kernel programming for GPU acceleration.

cuda-c-examples-torch

Compile and call CUDA C kernels from PyTorch using load_inline.

cuda-c-api

Reference CUDA C APIs for kernel launches and memory management.

cuda-c-optimization

Optimize CUDA C kernels for performance, stability, and debugging.

cuda-c-patterns

Provide CUDA-C kernel pattern templates for vector operations, reductions, and matrix multiplication.

pypto-case-elemwise-gelu

Generate a 1D GELU elementwise kernel with a tanh-free formula using PyPto.

pypto-case-matmul-2d

Generate tiled 2D matmul kernels with tail handling using the pypto DSL.

pypto-case-loss-crossentropy

Compute cross-entropy loss for batched Pypto kernels with two-stage tiling.

pypto-case-norm-batchnorm

Normalizes 3D tensors per-channel using Pypto BatchNorm-like operations.

Official

Intermediate

Frequently Asked Questions

FAQPage Schema

How to install AKG?▼

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

How to generate a GPU kernel automatically?▼

Describe the operator you need (for example a matmul or softmax), and the kernel-workflow skill builds a validated task, generates Triton or CUDA C code, and verifies it against the PyTorch reference.

Which hardware and DSLs does AKG support?▼

It supports NVIDIA GPUs (A100, V100), Ascend NPUs (910B, 310P, 950 series), and CPUs, with DSLs including Triton, CUDA C, TileLang, C++, and PyPTO.

Can AKG optimize an existing kernel for speed?▼

Yes. The adaptive_search and evolve workflows iteratively search tiling and scheduling parameters, then benchmark each candidate to find the fastest correct implementation.

Does AKG work with Claude Code and Cursor?▼

Yes. All skills follow the standard SKILL.md format and are symlinked for discovery in Claude Code, Cursor, and OpenCode.