How do I install the Areno Develop Kernel skill?

Run npx skillmds@latest add inclusionai/areno-develop-kernel in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Areno Develop Kernel safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Areno Develop Kernel?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Areno Develop Kernel free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Areno Develop Kernel?

inclusionai (@inclusionai) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Areno Develop Kernel by inclusionai

Name: Areno Develop Kernel
Author: inclusionai

Areno Develop Kernel

Develop, optimize, debug, and validate an AReno CUDA, Triton, fused, attention, convolution, routing, or MoE operator. Use when changes touch areno/accel or a runtime kernel boundary and require forward, backward, dtype, layout, CUDA graph, or benchmark validation.

inclusionai be7eb74 5 files · 7.4 KB Updated Sep 17, 2026

File contents

Develop an AReno Kernel

Define the mathematical reference, shapes, dtype, layout, supported devices, and backward contract before implementation.

Develop and commit on a local branch. A remote GPU checkout is validation-only: pull the committed branch there, and never edit or copy source files into it.

Generic harnesses

Both callables must accept one tensor and return one tensor:

python .agents/skills/areno-develop-kernel/scripts/check_operator.py \
  --reference package.module:reference --candidate package.module:candidate \
  --shape 8,16 --dtype float32 --device cuda

python .agents/skills/areno-develop-kernel/scripts/benchmark_operator.py \
  --callable package.module:candidate --shape 8,16 --dtype float16 --device cuda

For multi-input or stateful kernels, extend a focused test under tests/ rather than weakening this harness.

Workflow

Locate Python wrapper, extension registration, C++/CUDA source, engine layer, and model call site.
Add a small PyTorch reference and shape/dtype assertions.
Implement forward and backward before using it in training.
Test representative shapes, boundary shapes, supported non-contiguous layouts, finite output, and gradients.
Validate TP/sequence-parallel local shapes and CUDA graph capture/replay.
After pulling a branch that changes areno/accel, rebuild remotely with pip install -e . --no-deps --no-build-isolation. Do not reinstall for Python-only changes.
Benchmark only after correctness. Read references/kernel-checklist.md.

Do not add a silent production fallback for a required kernel. Report unsupported cases explicitly.

---name: areno-develop-kerneldescription: Develop, optimize, debug, and validate an AReno CUDA, Triton, fused, attention, convolution, routing, or MoE operator. Use when changes touch areno/accel or a runtime kernel boundary and require forward, backward, dtype, layout, CUDA graph, or benchmark validation.---# Develop an AReno KernelDefine the mathematical reference, shapes, dtype, layout, supported devices, and backward contract before implementation.Develop and commit on a local branch. A remote GPU checkout is validation-only:pull the committed branch there, and never edit or copy source files into it.## Generic harnessesBoth callables must accept one tensor and return one tensor:```bashpython .agents/skills/areno-develop-kernel/scripts/check_operator.py \ --reference package.module:reference --candidate package.module:candidate \ --shape 8,16 --dtype float32 --device cudapython .agents/skills/areno-develop-kernel/scripts/benchmark_operator.py \ --callable package.module:candidate --shape 8,16 --dtype float16 --device cuda```For multi-input or stateful kernels, extend a focused test under `tests/` rather than weakening this harness.## Workflow1. Locate Python wrapper, extension registration, C++/CUDA source, engine layer, and model call site.2. Add a small PyTorch reference and shape/dtype assertions.3. Implement forward and backward before using it in training.4. Test representative shapes, boundary shapes, supported non-contiguous layouts, finite output, and gradients.5. Validate TP/sequence-parallel local shapes and CUDA graph capture/replay.6. After pulling a branch that changes `areno/accel`, rebuild remotely with `pip install -e . --no-deps --no-build-isolation`. Do not reinstall for Python-only changes.7. Benchmark only after correctness. Read [references/kernel-checklist.md](references/kernel-checklist.md).Do not add a silent production fallback for a required kernel. Report unsupported cases explicitly.

Frequently asked questions