Triton-Ascend Agent Development Kit skill
24 skills and a reference for an agent writing Triton-Ascend kernels for Ascend NPUs: tiling, autotuning, accuracy checks
- 192 ★ on the repository
- Last commit
Minimal access
Why this level
- The kit consists of instructions, a reference and code templates; it runs nothing itself and contacts no services
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/triton-ascend-agent-development-kitDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- The kit consists of instructions, a reference and code templates; it runs nothing itself and contacts no services
This is third-party code. Review the repository files before installing.
What it does
The kit tunes an agent for Triton-Ascend kernel development, the Triton dialect for Ascend NPUs. AGENTS.md sets the working rules: forget the GPU and CUDA model, derive launch grid, tiling and alignment again from Ascend documentation, never invent signatures or supported types, and mark doubtful code with a comment that starts with ATTENTION!. The triton-ascend-dev-doc reference has 12 files, from quick start and the execution model to cube kernels, cube-vector fusion, autotuning, GPU-to-Ascend migration and an operator support matrix. The agent reads the triton-ascend-dev-doc.md index first and opens only the chapters its loading rules select. The skills folder holds 24 narrow task skills. They cover vector add, fused softmax, layer norm, fused attention, MoE routing, batched token reordering, preventing UB overflow, costmodel ranking, autotuning and accuracy validation against PyTorch with dtype-aware tolerances. Ready archives for Codex, OpenCode and Claude are published in the releases, and the build-workspace-agent-specific-kit.sh script produces the same archives.
Who it is for. Engineers who write kernels for Huawei Ascend NPUs with Claude Code, Codex or OpenCode and port GPU code to Triton-Ascend.
Good fit when
- You need to port a Triton kernel from GPU to Ascend and the agent must not carry over CUDA assumptions
- You are writing fused attention, softmax or layer norm for an NPU and want a proven tiling scheme
- Kernel autotuning on Ascend gives unstable results and you need a tuning order with costmodel ranking
- Kernel tests compare against PyTorch and tolerances for float16, bfloat16 and integers are picked by hand
- The agent gets lost in Ascend documentation and should load only the chapters for the current task
Not a fit when
- The kernels target NVIDIA or AMD: the kit tells the agent to drop GPU experience
- You want a general Triton course from scratch: the reference assumes you know the basics
- The bench has no Ascend NPU and no Triton-Ascend environment, so there is nothing to run the code on
Example request
Port this softmax kernel from GPU to Triton-Ascend, derive tiling from the Ascend documentation and check accuracy against PyTorchLimitations
- The author warns that Triton-Ascend and the NPUs themselves change quickly, so the agent's proposals may not fit your environment version. Check results against the documentation at hiascend.com and the triton-ascend repository.
- There is no plugin manager install: the archive from releases or the dist folder is unpacked into the agent's working folder by hand.
- The kit covers the listed 24 tasks. For other operators the agent relies on the reference and is required to ask clarifying questions.
- Skills and the reference are in English. The last commit in the repository is dated August 2026.
How to disable. Delete the unpacked files from the agent's working folder: AGENTS.md, triton-ascend-dev-doc.md, the triton-ascend-dev-doc folder and the skills from the skills folder.
FAQ
What is in the Triton-Ascend Agent Development Kit?
The kit holds AGENTS.md with rules, a 12-file reference, the triton-ascend-dev-doc.md index and 24 skills in the skills folder. Skills are split by task: fused attention, softmax, MoE, autotuning, accuracy validation and more.
How do I install it for Codex, OpenCode or Claude?
Download the archive for your agent from the repository releases and unpack it into the working folder. You can also build the archives with build-workspace-agent-specific-kit.sh, and the result lands in the dist folder.
Can I use the kit for NVIDIA kernels?
No, the kit targets only Ascend NPUs. The rules tell the agent to drop CUDA assumptions and derive launch and tiling again from Ascend documentation.
What does an ATTENTION! comment in the agent's code mean?
The agent uses it to mark a fragment it is unsure is correct for your Triton-Ascend version, hardware or API. A note beside it says exactly what to verify before running.
How does the kit check kernel accuracy?
The validate-triton-ascend-accuracy skill compares output with a PyTorch reference: shape and dtype must match, floating-point tensors use dtype-aware tolerances, integer and boolean tensors are compared exactly.
Discussion
No questions or reviews yet.
Related
SkillsOver 200 skills for development on Ascend NPU chips: device diagnostics, model conversion, training and profiling.
The author's personal set: shared agent rules and dozens of skills for GitHub, the browser, macOS, messaging and releases, plus a sync script
Softaworks Agent Toolkit
Softaworks Agent Skills
A marketplace of skills, subagents and commands for coding agents: development, documentation, planning and workplace communication
Over twenty skills for coding agents: MCP and hooks, plan mode, a bug-fix protocol, prompting and a bash safety hook
Installer of a working setup for Google Antigravity that puts rules, skills and workflows into a project with one npx command
Armin's personal toolkit for the Pi Coding Agent: 18 skills, extensions, commands and themes, with some skills portable to other agents
README in short
The README is short: it says the package is for an agent that helps write Triton-Ascend kernels and lists links to Ascend and Huawei documentation and the triton-ascend repositories. It then describes what is in the kit and how to get the archive.
README badge
Original SKILL.md
---
name: validate-triton-ascend-accuracy
description: Validate Triton-Ascend kernel outputs against PyTorch references with dtype-aware tolerances, exact integer checks, bfloat16 promotion, and boolean handling. Use when an agent writes or reviews NPU kernel tests, needs a reusable accuracy-comparison helper, or must choose an initial comparison policy for float16, bfloat16, float32, integer, or boolean outputs.
---
# Validate Triton-Ascend Accuracy
## Goal
Make correctness checks explicit and consistent across Triton-Ascend kernel tests.
## Workflow
1. Compute the PyTorch reference from the same inputs before interpreting performance.
2. Assert identical output shape and dtype.
3. Compare floating-point tensors with dtype-aware tolerances.
4. Compare integer and boolean tensors exactly.
5. Add edge shapes, tail masks, zeros, extreme values, and NaNs when the operator semantics allow them.
## Implementation Pattern
~~~python
def assert_ascend_close(actual, expected):
assert actual.shape == expected.shape
assert actual.dtype == expected.dtype
dtype = actual.dtype
if dtype == torch.float16:
torch.testing.assert_close(
actual, expected, rtol=1e-3, atol=1e-3, equal_nan=True
)
elif dtype == torch.bfloat16:
torch.testing.assert_close(
actual.float(), expected.float(),
rtol=1e-3, atol=1e-3, equal_nan=True,
)
elif dtype == torch.float32:
torch.testing.assert_close(
actual, expected, rtol=1e-4, atol=1e-4, equal_nan=True
)