291 skills · sorted by relevance
Systematically identify whether a GPU kernel is compute-bound, memory-bound, or latency-bound using roofline analysis, occupancy calculations, compute/load ratio per tile, and SASS instruction inspection. Produces a decision matrix for optimization strategy selection (cp.async, warp interleaving, tiling, double-buffering, or CuAssembler hand-tuning).
The English summary is being prepared. Showing the original description.
Use when writing CUDA kernels, managing the thread, block, and grid hierarchy, tiling shared memory, using streams, setting nvcc flags, or using Thrust. Not for kernel debugging: use cuda-debugging.
The English summary is being prepared. Showing the original description.
Use when writing HIP kernels with hipcc, porting CUDA code through HIPIFY, profiling with rocprofv3, debugging with rocgdb, or optimizing for MI300X. Not for NVIDIA-side kernels: use cuda.
The English summary is being prepared. Showing the original description.
Use when choosing kernel spinlocks versus mutexes, using RCU, seqlocks, completions, or memory barriers, or debugging scheduling-while-atomic. Not for userspace atomics: use concurrency-debugging.
The English summary is being prepared. Showing the original description.
Use when debugging the Linux kernel: kgdb and kdb, ftrace and kprobes, dynamic debug, kdump and crash analysis, or printk levels on a live or crashed target.
The English summary is being prepared. Showing the original description.
Use when diagnosing scheduler latency, kmalloc vs vmalloc, page cache, meminfo under pressure, or OOM victim choice, or when reading kernel/sched, mm, or fs source.
The English summary is being prepared. Showing the original description.
Use when using kmalloc, vmalloc, or the page allocator, sizing kmalloc allocations, working with SLUB or the buddy allocator, or debugging memory zones and kernel OOM.
The English summary is being prepared. Showing the original description.
Use when writing KUnit tests, adding kselftest cases, fuzzing syscalls with syzkaller and kcov, running LTP, or wiring CI for a kernel patch.
The English summary is being prepared. Showing the original description.
Use when navigating kernel source, understanding boot flow, initcall levels, or major subsystems (VFS, scheduler, MM) in linux.git. Not for driver-model depth: use device-drivers.
The English summary is being prepared. Showing the original description.
Use when writing loadable kernel modules: Kbuild, module parameters, proc and sysfs entries, char devices, or ftrace debugging. Not for driver architecture: use device-drivers.
The English summary is being prepared. Showing the original description.
Post-exploitation workflow. Covers privilege escalation (Linux SUID/sudo/kernel, Windows UAC/service/token), persistence assessment, local enumeration, credential harvesting, and pivot preparation. Structured workflows for Linux and Windows targets using impacket, netexec, john, linpeas/winpeas, and standard Kali tools. Includes kernel exploit reference tables, GTFOBins exploitation chains, Potato attack selection, Docker/container escapes, DLL hijacking, SSH key harvesting, credential recovery from memory, and Windows token manipulation. Chains from /pentester or /credential-audit when access is obtained.
The English summary is being prepared. Showing the original description.
Reference for MCP Jupyter tools (kernel management, cell execution, Papermill). Use when executing notebooks, managing kernels, or running code interactively via MCP.
The English summary is being prepared. Showing the original description.
融合单元的范围判定与收益前置评估：判定「哪些计算该被融进同一个融合单元、边界画在哪、 融了值不值」。结构侧按融合单元构造规则切边界——函数边界即候选组；FA（含 BSA 等变种）为 禁融锚点；以 norm / rope / FA 为分界分区，全 Vec 区域优先试融；区域内有 Cube 时以 Cube 为 首算子向后包裹 Vec（Σvec 小于 cube 或遇下一 Cube 封口）；一轮之后纯 Vector 单元之间再迭代； MLP / MoE 以矩阵乘族（MatMul / BatchMatMul / GroupedMatmul）为界；若该段计算已由单个 融合算子承载，按规则 6 先确认载体是否已生效（不重复建单元）。 收益侧按计算单元利用率判型（memory_bound、*_vec|mac|mte2|mte3_ratio、cube_utilization）， 叠加带宽下限 / 区域占比 / 天花板法 / 地板先行 / Amdahl 传导校验给出 go/no-go。 当用户问「这几个算子能不能融成一个」「融合边界画在哪」「融多大范围」「融了值不值得」， 或 profiling-analyze 给出融合机会候选需要定范围与收益时使用；即使用户只说「这条链要不要 合成一个 kernel」「这里能不能省一次读写」也应触发。 **已经做完、实测收益却打平或为负时同样触发**——此时的任务不是接受"融合无收益"， 而是按事后复核做**等功忙计数对照**归因，并在缺陷落在融合核自身时**升级 operator-dev**。 边界：只出「融合单元草案 + 收益判定 + 否决理由」。不做采纳与选点（dit-perf-opt 依收益选 top 并派活）、不做采集与瓶颈定位（profiling-collect / profiling-analyze）、不写 kernel、不写 pattern、 不做使能与并行选型、不定义验收标准（引 accuracy-gate / perf-gate）。 由 profiling-analyze（融合机会候选）与 operator-dev（实现前的收益评估）调用。
The English summary is being prepared. Showing the original description.
算子级开发与性能优化：Triton / Ascend C / Catlass / PyPTO / TileLang 算子 编写、精度对齐与性能调优。优先路由到外部 cannbot-skills 技能库（不重复其内容）， 本 skill 只保留场景 → skill 映射与 MindIE-SD 特有补充。 当用户需要新增/优化算子、定位算子性能或精度问题时使用此 skill。 即使用户只提到"写个 triton kernel"或"这个算子怎么加速"而未说算子，也应触发； pattern 融合/编译后端见 pattern-dev，算子基准选型/接入测试见 benchmark-dev。 由 dev-workflow 或 pattern-dev 的 replacement kernel 场景触发。
The English summary is being prepared. Showing the original description.
**精度验收标准**：凡是"改动不应改变结果"的场合（等价替换、算子/子模块融合、并行切分、 编译与图下发、拷贝消减），都按本标准判"合格 / 不合格"——等价分层（L1 逐位 / L2 数值门 / L3 有损门）+ 三级验收序（① 同配置重跑逐位 → ② 跨配置数值门 + 产物 md5 不变 → ③ 质量门）+ 判据不达标时的排障入口。当用户说"这个算子能不能换个写法/换个 kernel/等价实现/ 无损替换""替换后结果会不会变""怎么证明逐位一致""结果不对""花屏""尾部塌了" "CPU 跑对 NPU 跑不对"，或发现某个 hotspot 占了大头（例如某类算子在阶段里占 90% 以上）想动手时， 都应触发；即使用户只说"这样改有没有把结果改坏""能不能判它是无损的"也应触发。 覆盖：等价分层与判据、满足本标准的实现约定（per-shape 对拍写进实现，把索引/相位/顺序类重写 错误在首次调用抓住）、per-shape 条件性等价（同一改动在不同形状下结论可能不同）、 判据不达标 → `references/silent-failure-localization.md` 排障入口、否决案例的形态学 （换写法未换 kernel / 差异极小仍非逐位 / 等价但 OOM）、以及收益口径的归属（性能数字一律交 `perf-gate`）。 **近义分流**：选量化档 / 特性选档（要不要开量化、开哪一档）→ `dit-perf-opt`； 并行选型（USP / CP / TP 怎么切）→ `dit-parallel-opt`；本技能只判"结果是否被改变"， 不负责选档与选型。
The English summary is being prepared. Showing the original description.
对 profiling 数据（kernel_details.csv / trace_view.json / step_trace_time.csv）做统一分析： 5 层递进管道（analyze_trace.py）+ baseline/优化后 kernel diff（compare_traces.py），输出 瓶颈诊断、算子执行序、可融合候选与 P0-P2 方向建议。输入统一来自 profiling-collect。 只要用户有 profiling 产出并问"为什么慢/瓶颈在哪/优化前后差多少/往哪个方向"，都用本技能；采集走 profiling-collect，方案选档走 dit-perf-opt，实现/并行/基准走 pattern-dev/dit-parallel-opt/benchmark-dev——本技能只做诊断与方向。 由 model-auto-optimization 的 S1（融合分析）/S3（并行）阶段调用，亦由 dev-workflow 的分析阶段指引加载。 融合机会候选只做**识别**；候选的**边界与收益判定**（能不能融成一个单元、融多大、值不值）交 `fusion-scope-analyze`（本技能的辅助技能）——本技能不判融合边界。
The English summary is being prepared. Showing the original description.
Convert PyTorch AT_DISPATCH macros to AT_DISPATCH_V2 format in ATen C++ code. Use when porting AT_DISPATCH_ALL_TYPES_AND*, AT_DISPATCH_FLOATING_TYPES*, or other dispatch macros to the new v2 API. For ATen kernel files, CUDA kernels, and native operator implementations.
The English summary is being prepared. Showing the original description.
Iterative Python via live Jupyter kernel (hamelnb).
The English summary is being prepared. Showing the original description.
Use when developing, reviewing, debugging, porting, or engineering Linux kernel device drivers, kernel modules, device tree bindings, bus clients, subsystem drivers, DMA/IRQ/memory paths, userspace interfaces, or hardware bring-up plans.
The English summary is being prepared. Showing the original description.
Distill an already-built or near-launch product into one incisive brand kernel — the "reveal" already hiding in its name, story, or the product itself — then build the name, voice, visual identity, and a developer-ready handoff kit around it. This is LATE-STAGE branding: use it when there is a real, thought-through product that is about to launch, or already live but not explained in a way people grasp in an instant. Use whenever the user wants to name or rename a product, find a tagline or slogan, create a brand identity / brand kit / style guide / design system, design a logo, wordmark, palette, or type system, mock up a homepage for a launch, or hand brand assets to a coding agent (Claude Code, Codex, Cursor). Trigger even when the user only says their product needs to "land" or "feel like" something, or that they can't explain it simply yet. Do NOT use for half-baked, pre-product ideas with nothing to excavate — this skill extracts a kernel that already exists; it does not invent one from a blank canvas.
The English summary is being prepared. Showing the original description.
Implement eBPF-based runtime security observability and enforcement in Kubernetes clusters using Cilium Tetragon for kernel-level threat Tespit and policy enforcement.
The English summary is being prepared. Showing the original description.
Iterative Python via live Jupyter kernel (hamelnb).
The English summary is being prepared. Showing the original description.
Performs Linux memory acquisition using LiME (Linux Memory Extractor) kernel module and analysis with Volatility 3 framework. Extracts process lists, network connections, bash history, loaded kernel modules, and injected code from Linux memory images. Use when performing incident response on compromised Linux systems.
The English summary is being prepared. Showing the original description.
Apple グラフィックス。Metal, Core Animation, Core Graphics, Core Image, SpriteKit, SceneKit。 MTLDevice, MTKView, MTLBuffer, MTLTexture, MTLResidencySet, MTLTensor, Metal 4 (MTL4CommandQueue, MTL4CommandBuffer, MTL4CommandAllocator, MTL4ArgumentTable, MTL4RenderCommandEncoder, MTL4Compiler), CALayer, CABasicAnimation, CADisplayLink, CAMetalDisplayLink, CAMetalDrawable, CGContext, CGPath, CGImage, CIFilter, CIContext, CIImage, CIRAWFilter, CIWarpKernel, CIBlendKernel, SKScene, SKSpriteNode, SKAction, SCNNode, SCNGeometry, SCNMaterial。
The English summary is being prepared. Showing the original description.