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-ascend-hardware-constraintsgit 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-ascend-hardware-constraints)<a href="https://agentmods.dev/skills/mindspore-ai/akg/triton-ascend-ascend-hardware-constraints"><img src="https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-ascend-ascend-hardware-constraints/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-ascend-hardware-constraints"><img src="https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-ascend-ascend-hardware-constraints.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 | $0.00079 | $0.01205 |
| Opus 5 | $0.00039 | $0.00602 |
| Sonnet 5 | $0.00016 | $0.00241 |
| Haiku 4.5 | $0.00008 | $0.00120 |
Grade A, and why
triton-ascend-ascend-hardware-constraints 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 9d 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.
How it starts
The opening of the file, as written. The whole thing — 92 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Ascend 硬件约束与编译器限制
不同型号的硬件具体容量不同,算子生成时会同步传入硬件信息文档,以下公式中的容量值请参考该文档。
1. 存储层级预算
CUBE 路径(matmul / tl.dot)
Matmul 数据走 L0A/L0B/L0C,不经过 UB:
| 缓冲区 | 用途 | 约束公式 |
|---|---|---|
| L0A | 左矩阵 A tile (m0 × k0) | m0 × k0 × sizeof(A.dtype) ≤ L0A容量 |
| L0B | 右矩阵 B tile (k0 × n0) | k0 × n0 × sizeof(B.dtype) ≤ L0B容量 |
| L0C | 结果 C tile (m0 × n0),支持累加 | m0 × n0 × sizeof(C.dtype) ≤ L0C容量 |
计算示例(以某硬件 L0A = 64KB 为例):
- fp16(2 字节/元素):可容纳 32K 个元素 → BLOCK_M=128, BLOCK_K=256 恰好填满
- fp32(4 字节/元素):可容纳 16K 个元素 → BLOCK_M=128, BLOCK_K=128 恰好填满
选择 tile 尺寸时,确保三个缓冲区都不超限。fp32 占用是 fp16 的 2 倍,需相应缩小 tile。
VEC 路径(element-wise / reduce / norm)
向量运算数据走 UB:
| 缓冲区 | 用途 | 约束公式 |
|---|---|---|
| UB | 所有活跃 tensor 和中间变量 | BLOCK_SIZE × sizeof(dtype) × 活跃tensor数 × multi_buffer系数 ≤ UB容量 |
编译器启用 auto-multi-buffer 后,实际占用约为基础量的 2~3 倍。kernel 中的中间变量(如 tl.where 产生的临时缓冲)也占用 UB,实际占用会显著高于 BLOCK_SIZE × sizeof(dtype) × 输入数。
tile 选择策略:从较大 BLOCK_SIZE 开始尝试,遇到 ub overflow 编译错误时逐级缩小。
2. bishengIR 编译器已知限制
2.1 range() 边界不可混用运行时变量
# 编译器崩溃(bishengIR SIGABRT)
for k in range(start_n, start_m + BLOCK, BLOCK_K):
...
start_n、start_m 是运行时值,BLOCK、BLOCK_K 是 tl.constexpr。这种混合用法会导致编译器内部错误。
规避方案:使用全 constexpr 的 range,在循环体内用运行时 if 跳过无效迭代:
for k in range(0, N, BLOCK_K): # N 和 BLOCK_K 都是 constexpr
# 可选:运行时条件跳过无效块
...
2.2 复杂 mask + tl.where 导致 HiVM 错误
当嵌套 mask 组合传入 tl.where 时,编译器后端可能报 hivm.hir.vsel: Unsupported op for finding the root alloc。
规避方案:用乘法替代 tl.where,将 bool mask 转为 float 后与数据相乘:
# 触发 hivm.hir.vsel 错误
a = tl.where(tri_mask & bounds_mask, a, 0.0)
# 规避:mask 转 float 后相乘
a = a * tri_mask.to(tl.float16) * bounds_mask.to(tl.float16)
2.3 其他编译器限制
详见 debugging 文档中的「禁止使用的语法」完整列表。
3. Strided memory access 的性能代价
Ascend 硬件对非连续内存访问有显著性能惩罚。当 kernel 的核心路径包含 stride > 1 的内存访问模式(如 pooling 的滑窗、dilated convolution 的间隔采样),Triton 生成的代码需要逐元素或小块 gather,而 CANN 原生算子可能使用硬件数据搬运单元(MTE)的专用模式,性能差距可达数十倍。
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.
- 9d ago First seen · 92 lines · 79 tokens per session scan A 3f1edab7057d
triton-ascend-ascend-hardware-constraints is a skill published in the GitHub repository mindspore-ai/akg (259 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 79 tokens to every session and 1,205 once invoked, about $0.0004 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
spark-environment-setup
Set up a working ML training/inference environment on NVIDIA DGX Spark (GB10, aarch64, CUDA 13). Use when installing PyTorch/Unsloth/TRL/vLLM on DGX Spark, hitting libcudart or wheel-ABI errors on aarch64, or choosing between NGC containers and bare pip installs.
spark-memory-thermal-ops
Manage unified memory and thermals during long-running ML jobs on NVIDIA DGX Spark. Use when planning memory headroom for a training run on GB10, when a job OOMs on unified memory, or when monitoring temperature and power during multi-hour training.
spark-training-gotchas
Preflight and diagnose the ten known failure modes for ML training on NVIDIA DGX Spark. Use when a training run on DGX Spark fails to start, OOMs below the 128GB limit, slows down mid-run, or before any multi-hour training job on GB10.
llama-cpp
Runs LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA hardware. Use for edge deployment, M1/M2/M3 Macs, AMD/Intel GPUs, or when CUDA is unavailable. Supports GGUF quantization (1.5-8 bit) for reduced memory and 4-10× speedup vs PyTorch on CPU.
minicpm5-deploy-vllm-ascend
Deploy MiniCPM5-2B with vLLM on Huawei Ascend NPU using vLLM-Ascend. Use when the user mentions vLLM-Ascend, Ascend NPU, Huawei Ascend, CANN, torchnpu, davinci devices, or wants an OpenAI-compatible MiniCPM5 server on Ascend hardware.
minicpm5-deploy-litert
Run MiniCPM5-2B or MiniCPM5-1B on-device with Google's LiteRT-LM runtime — the litert-lm CLI or its OpenAI-compatible server on a desktop, the Kotlin API or the AI Edge Gallery app on Android, the same .litertlm bundle on CPU or GPU. Use when the user says "LiteRT", "LiteRT-LM", "litertlm", ".litertlm", "Android"…