Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add mindspore-ai/akg --skill triton-ascend-optimizationgit clone --depth 1 https://github.com/mindspore-ai/akgWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/mindspore-ai/akg/triton-ascend-optimization)<a href="https://agentmods.dev/skills/mindspore-ai/akg/triton-ascend-optimization"><img src="https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-ascend-optimization/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/mindspore-ai/akg/triton-ascend-optimization"><img src="https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-ascend-optimization.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00146 | $0.00991 |
| Opus 5 | $0.00073 | $0.00495 |
| Sonnet 5 | $0.00029 | $0.00198 |
| Haiku 4.5 | $0.00015 | $0.00099 |
Grade A, and why
triton-ascend-optimization scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 6d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
What it actually says
Triton Ascend 性能优化指南
优化策略 Checklist
- Grid 1D 化:
grid=(CORE_NUM,)+ 核内交错循环for block_id in range(pid, total, CORE_NUM) - Grid 维度选择:
- 对于计算密集型算子(矩阵乘、卷积、大块 reduce等),考虑 2D/3D grid,利用硬件调度优势
- 对于大量小规模计算(element-wise、pointwise等),考虑 1D grid + 核内循环,减少启动开销
- 核内循环: 无需 for 的场景添加额外循环,编译器自动多级流水
- 尝试不同 BLOCK_SIZE: 从较大 tile 开始,
ub overflow则缩小;在核内循环中平衡并行度和资源占用:- 尝试小切分策略,使读写能够并行进行
- 尝试大切分策略,提升 UB 使用率
- 列多组参数配置,添加 @triton.autotune
- 算子拆分: 复杂融合算子可拆为多 kernel 顺序执行,有时性能更优
- Autotune: 列多组 tile 参数配置(不含 num_warps/num_stages)
- Reduction 用标量累加: 每个核心标量累加 + 单次 atomic 写入
- 内存对齐: matmul 的 K 维度按 512B 对齐提升带宽
- 避免 host 端 permute: 非最后维 reduce 在 kernel 内用多维索引处理
- 隐式广播: 用
a[:, None] * b替代tl.broadcast_to,减少临时 tensor - load 时直接 mask:
tl.load(ptr, mask=m, other=0.0)优于先加载再tl.where - 减少冗余精度转换: 避免反复在 fp16/fp32 转换 即
.to(float16)和.to(float32),一次转换多次复用 - 核心数配置: grid 数设为核心数(VEC/CUBE),过大时启动开销反增
- 256B 对齐: 数据搬运以 256B 为单位,对齐可提升带宽
Reduction 优化
每个核心先局部标量累加,最后一次原子写入:
core_sum = 0.0
for block_start in range(pid, total_blocks, CORE_NUM):
data = tl.load(...)
core_sum += tl.sum(data, axis=0)
tl.atomic_add(output_ptr, core_sum)
数值稳定性
防溢出
max_val = tl.max(scores, axis=0)
scores = scores - max_val
p = tl.math.exp2(scores)
防负值开方
- 任何 sqrt 前确保非负:
max(input, 0.)或max(input, eps)
精度提升
- matmul 使用 fp32 累加器:
acc = tl.zeros([M, N], dtype=tl.float32) - 最后再转回目标精度:
result = acc.to(tl.float16)
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 6d ago First seen · 68 lines · 146 tokens per session scan A 021f73e20037
triton-ascend-optimization is a skill published in the GitHub repository mindspore-ai/akg (259 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 146 tokens to every session and 991 once invoked, about $0.0007 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-03.
Other skills, from other repositories
ruview-quickstart
Onboarding and first-run for RuView (WiFi-DensePose) — Docker demo with simulated data, repo build, and the fastest path to a live sensing dashboard. Use when someone is new to RuView or wants the shortest path to "it works on my machine".
onboard
Zero-to-sensing path picker for RuView (WiFi-DensePose) — pick docker-demo, repo-build, or live-esp32 and run the next concrete step.
lesson-quiz
Test a learner on a single Claude Code tutorial lesson (01-10) with 10 questions, scoring answers and flagging weak spots. Use before, during, or after a lesson. Don't use for whole-tutorial assessment or explaining a topic instead of testing it.
doca-programming-guide
Use this skill when the user is writing their first DOCA app or asking a library-agnostic programming question — picking a shipped sample to copy and modify, wiring the canonical pkg-config doca-{library} + meson build (or FFI from Rust / Go / Python against the public C ABI), walking the cfg-create → init → start →…
devin-akin-perspective
A vendor-neutral perspective for making enterprise Wi-Fi decisions, with emphasis on radio physics, training, technical standards, and checking vendor claims against real deployments.
design-logic-circuit
Design combinational logic circuits from a functional specification through gate-level implementation. Covers AND, OR, NOT, XOR, NAND, NOR gates; NAND/NOR universality conversions; and standard building blocks including multiplexers, decoders, half/full adders, and ripple-carry adders. Use when translating a Boolean…