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-debugginggit 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-debugging)<a href="https://agentmods.dev/skills/mindspore-ai/akg/triton-ascend-debugging"><img src="https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-ascend-debugging/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-debugging"><img src="https://agentmods.dev/badge/skills/mindspore-ai/akg/triton-ascend-debugging.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 | $0.00069 | $0.01583 |
| Opus 5 | $0.00034 | $0.00792 |
| Sonnet 5 | $0.00014 | $0.00317 |
| Haiku 4.5 | $0.00007 | $0.00158 |
Grade A, and why
triton-ascend-debugging 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.
What it actually says
调试与排查清单
完整调试清单
内存访问问题
- 所有 load/store 是否都有 mask 或 boundary_check？
- stride 参数设置是否正确？
- 数组索引是否越界？
控制流检查
- 是否误用了 return/break/continue？
- 复杂条件是否用 mask 组合实现？
-
tl.constexpr是否只在内核参数中使用？
Grid 与 Block 配置检查
- Grid 总大小是否不超过 65535？
- 对于大 shape 算子，是否采用了交错循环
for i in range(pid, total, core_num)？ - Grid 维度是否为 tuple 类型且不超过 3 维？
并发与原子操作检查
- 并发写入是否使用了原子操作（
tl.atomic_add等）？ - 原子操作是否必要（能否避免）？
切片与索引检查
- 是否避免了Python风格的直接切片（如
b[0]、b[i:j]）？ - 是否对
tl.arange生成的张量误用了tl.get_element？ - 切片操作是否使用了正确的API（
tl.get_element、tl.extract_slice等）？
性能优化检查
- 内存访问是否连续（避免跨步访问）？
- 是否充分利用了块内并行？
- 复杂算子是否考虑拆分为多个简单kernel？
禁止使用的语法（Ascend 后端）
| 禁止写法 | 替代方案 |
|---|---|
return / break / continue |
使用 mask 控制流程 |
| lambda 表达式 | 内联函数或 tl.where |
链式布尔运算 a and b |
分步计算 mask：m1 = ...; m2 = ...; m = m1 & m2 |
张量直接索引 tensor[i] |
tl.load(ptr + offset) / tl.store(ptr + offset, val) |
Python 切片 b[0] / b[i:j] |
tl.get_element / tl.extract_slice / tl.insert_slice |
对 tl.arange 结果用 get_element |
直接计算索引值 |
while 循环 |
for i in range(MAX): if i < n: |
range() 混用运行时变量和 constexpr |
全 constexpr 的 range(0, N, BLOCK_K) + 循环体内运行时 if |
tl.float16(scalar) |
scalar.to(tl.float16) |
tl.constexpr 在 host 侧使用 |
仅在 kernel 参数中使用 |
| if-else 中负偏移 | tl.maximum(offset, 0) |
复杂 tl.where 用于内存偏移 |
拆分为 if-else 静态分支 |
常见错误速查表
编译错误
| 错误类型 | 典型症状 | 常见原因 | 解决方案 |
|---|---|---|---|
| UB/CBUF 溢出 | ub overflow, requires X bits while 1572864 bits available |
BLOCK 尺寸过大或中间变量过多 | 缩小 BLOCK 尺寸；减少同时活跃的 tensor 数 |
| HiVM vsel 错误 | hivm.hir.vsel: Unsupported op for finding the root alloc |
嵌套 mask + tl.where 组合过于复杂 | 用乘法替代 tl.where：a * mask.to(dtype) |
| 内存越界访问 | 运行时错误、结果异常、随机崩溃 | load/store缺少mask或boundary_check | 添加正确的mask或boundary_check保护 |
| Grid超限 | 编译失败或运行时错误 | grid总大小超过65535 | 使用交错循环for i in range(pid, total, core_num)或连续分块处理 |
| 控制流错误 | unsupported AST node type: Continue |
使用了return/break/continue | 改用 if-else 包裹逻辑 |
| while循环错误 | 编译失败（Ascend后端） | 使用了while循环 | 改用for + if替代：for i in range(MAX): if i < n: |
| constexpr 索引 | ValueError('unsupported tensor index: constexpr[0]') |
对 tl.sum 等返回的标量做 [0] 索引 |
直接使用标量结果，不要索引 |
| 切片语法错误 | 编译失败 | 使用了b[0]或b[i:j]直接切片 |
使用tl.get_element或tl.extract_slice |
| tl.arange索引错误 | 编译失败 | 对tl.arange结果使用get_element |
直接计算索引值而非提取 |
| 类型转换错误 | cast incompatible |
隐式 cast 或使用tl.float16(scalar) |
用 .to(tl.float16) 或显式指定 acc dtype |
| constexpr误用 | 编译失败 | 在host侧使用tl.constexpr | 仅在kernel参数中使用tl.constexpr |
| Stride设置错误 | 计算结果错误、数据错位 | stride参数计算或传递错误 | 验证stride设置，检查tensor.stride() |
| 数值不稳定 | 结果为NaN或Inf | softmax/sqrt等操作溢出 | 减去最大值、检查非负、使用float32 |
| 数据竞争 | 结果不确定、每次运行不同 | 多program并发写入同一位置 | 使用tl.atomic_add等原子操作 |
| BLOCK_SIZE过大 | 编译失败或运行时错误 | BLOCK_SIZE超过65536或硬件限制 | 减小BLOCK_SIZE，使用循环处理 |
| tl.where偏移计算 | 编译失败（Ascend后端） | 在内存偏移中使用tl.where | 改用if-else静态分支处理 |
| 性能低下 | 运行缓慢 | 内存访问不连续、切分不合理 | 优化内存布局、调整BLOCK_SIZE、使用block_ptr |
| 运行时range边界崩溃 | bishengIR crash | range()的start/stop混用运行时变量和constexpr | 改用全constexpr的range(0, N, BLOCK_K)，循环体内用运行时if跳过 |
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 · 85 lines · 69 tokens per session scan A 4b00bc1d6486
triton-ascend-debugging is a skill published in the GitHub repository mindspore-ai/akg (259 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 69 tokens to every session and 1,583 once invoked, about $0.0003 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-optimization
Optimize Apache Spark jobs with partitioning, caching, shuffle optimization, and memory tuning. Use when improving Spark performance, debugging slow jobs, or scaling data processing pipelines.
notebooklm
Install, authenticate, troubleshoot, and operate Gemini Notebook through the notebooklm-py CLI or typed async Python API. Use for notebook and source management, grounded chat and research, and artifact generation or download when the user mentions Gemini Notebook, notebooklm-py, the notebooklm CLI, or its Python API.…
debug-inference
Debug inference clients that use an attached provider and its native endpoint, including hosted APIs and host-local Ollama, vLLM, SGLang, TRT-LLM, LM Studio, or NIM. Use for provider attachment, endpoint policy, credential substitution, topology, and migration from the removed inference.local endpoint. Trigger…
oh-my-posh
Install, configure, or troubleshoot Oh My Posh/ohmyposh: shell init, themes, segments, Nerd Font icons, and prompt setup on PowerShell, zsh, bash, or fish.
eagle3-triage
Triage a failed EAGLE3 pipeline run. Identifies which step failed (data synthesis, hidden state dump, training, or benchmark), diagnoses root cause from logs, and suggests fixes. Use when user reports an EAGLE3 pipeline failure or asks why a specific step failed. Also helps debug new model support issues.
trulens-instrumentation
Instrument LLM apps with TruLens OTEL-based tracing - from setup to debugging and optimization.