Showing 3961–3990 of 10205 skills
多卡分布式训练中的 loss/gnorm 精度溢出检测与根因追溯。基于 MSProbe dump 数据,先跨 rank 定位首次出现 NaN 的源卡,再在源卡上追溯具体的溢出根因算子。 当用户需要:(1) 多卡分布式训练场景下的 NaN/Inf 溢出检测 (2) 找出首先出现 NaN 的源卡 (3) 追溯根因计算算子 (4) loss/gnorm NaN 问题定位 时使用此 skill。
kali20gakki/msAgent
通用训推一致性(prefill)的 dump 数据分析 skill。分析**已采集**的训推两侧 dump:探查 dump 结构 → 语义边界映射(不要求算子同名)→ 逐边界精确门禁比对 → 定位首差异 → ULP 量化区分舍入/逻辑级 → 追放大链 → 差异点隔离验证(打桩迭代:输出打桩指令给采集侧,等它用 msprobe `load` 覆盖边界再采后比下游,枚举所有独立差异点)→ 排除根因假设 → 输出整体结论 → 产出分析报告。与具体框架/硬件解耦,适用 msprobe PrecisionDebugger 或等价工具采的任何 dump。触发词:训推一致性分析、dump 数据分析、dump 已采好帮我分析、首差异定位、所有差异点、差异点隔离、打桩指令、对齐报告、训推不一致结论、比较两侧 tensor。采集与打桩再采由 rl-prefill-consistency-dump 负责,本 skill 只消费已落盘 dump、输出打桩指令驱动再采。
kali20gakki/msAgent
训推一致性(prefill)采集侧:为训练/推理两侧采集可比的 prefill dump(输入对齐 PROMPTS_ONLY → 参数对齐校验(dump 前前置:校验两侧 config 超参/结构/权重,不一致输出参数不一致点并对齐)→ msprobe 接入采集 → 采集验证 → 恢复代码)。按用户需求判断用法:① 仅采集 dump——用户只要 dump 数据,走采集主流程(0→7→9),直接产出两侧可比 dump;② 配合分析——用户要求找所有独立差异点/差异点隔离验证/配合 rl-consistency-dump-analysis 分析时,才按分析侧打桩指令用 msprobe `load` 覆盖边界二次采集(阶段 8)再恢复代码。差异分析、首差异定位、打桩指令、整体结论由 rl-consistency-dump-analysis 承担。触发词:训推一致、训推一致性、prefill 比对、msprobe dump/采集、精度比对数据采集、打桩二次采集、参数对齐校验、参数不一致点、mHC 配置不一致、权重不同步、清卡验环境后采数据;已采 dump 的分析转 rl-consistency-dump-analysis。
kali20gakki/msAgent
RL 强化学习训练与推理(rollout)路径的算子差异性扫描。当用户提供训练启动脚本路径,要求对比 Megatron 训练路径与 vLLM 推理路径的算子差异(尤其是融合算子 vs 单算子不一致)时使用此 skill。典型触发词包括"训推算子扫描"、"训推差异性"、"算子差异报告"、"融合算子对比"、"train vs infer op diff"。注意:本 skill 通过运行完整 RL 训练脚本(集成 profiling)来采集真实运行时算子(非静态源码扫描,非独立分离运行)。当训练、推理使用不同脚本时禁止触发。
kali20gakki/msAgent
当用户直接要求进行数据集压缩测评,或在量化调优过程中需要使用 AISBench 评测工具输出数据集精度等场景时使用。 本 Skill 负责:使用 AISBench 数据集压缩代码、用 RBF Kernel Herding 算法生成压缩后的数据子集(coreset),并按用户目标交付两种产物之一: ① 接入量化调优主流程:产出独立的 coreset 数据集配置(不覆盖原始数据),回传「全集 config_name + coreset config_name」给编排层,供 evaluation.yaml 切换; ② 仅数据集压缩测评(不做量化调优):基于 example 模板生成「全集测试脚本」和「子集测试脚本」两份 shell 脚本。 暂时支持 AIME 2025 和 GPQA,其他数据集需要再结合 aisbench tools目录下 herding_coreset_selector 源码进行分析。
kali20gakki/msAgent
Analyze Ascend NPU computation-side profiling data for single-card runs or a selected rank from multi-card runs. Use this skill when the user asks to diagnose computation bottlenecks, AI Core / AI Vector / AICPU hotspots, dynamic shape overhead, block dim issues, redundant TransData/Transpose/Cast, cross-stream waits, frequency/runtime-state issues, or fusion opportunities.
kali20gakki/msAgent
针对昇腾 NPU 上的 PyTorch 模型,识别可做融合算子替换的代码段、替换为 torch_npu 融合算子并做性能/精度验证。当用户提到融合算子、NPU 算子替换、RMSNorm/SwiGLU/RoPE/Attention 融合、torch_npu 亲和算子等关键词是,使用本 Skill。
kali20gakki/msAgent
Adapt torch_npu.profiler interface for AI frameworks and training/inference code on Ascend NPU. Use this skill whenever users need to integrate profiling, add performance data collection, or instrument PyTorch code with torch_npu.profiler on Huawei Ascend NPU platforms. Triggers on: "torch_npu profiler", "NPU profiling", "Ascend profiler", "profiler adaptation", "profile training", "profile inference", "add profiler to framework", "性能采集", "profiling 数据", "昇腾 profiler", and any request to add performance monitoring to Ascend NPU code.
kali20gakki/msAgent
在 W8A8 动态量化后,用 MindIE-SD 冒烟测试 DiT 量化权重是否可加载(可选 skill)。与 `quant-tuning-evaluate`(DiT 扩展节)的区别:本 skill 仅做"能跑通 + 权重可用"的轻量验证(4~16 条 prompt,几分钟出结果),不算分;后者产出视频集交 `quant-tuning-score-dit`(AISBench-VBench 16 维度)评分,是调优回路一环。不引入 vLLM-Ascend、不做精度指标评估(FID/CLIP-score 等)。
kali20gakki/msAgent
量化专家经验调优库:按「L1 通用专家调优意见 → L2 结构化专家调优意见 → L3 模型专属专家调优意见」三级递进, 输出量化精度调优的完整手段(离群值抑制 / 量化方法·粒度·对称性选型 / 校准集调整 / 敏感层回退 / 模型专属策略), 每条结论附专家原因分析与专家意见可信度等级。 本 Skill 是专家经验库,负责说明可选离群值抑制算法并询问用户选择哪些算法,以及回答「怎么调、哪些层需要回退及为什么」;不运行离群值抑制 processor,不修改 YAML、不执行量化、不做 EP 检查 / 服务化 / 任务评测。
kali20gakki/msAgent
Use after a skill or governance-surface change to regenerate the control surfaces and skill mirrors.
tvproductions/gzkit
Use when the operator asks for a "diet" or "progressive disclosure" pass on AGENTS.md, CLAUDE.md or the rules.
tvproductions/gzkit
Use when inspecting or operating Herdr sessions, workspaces, tabs, panes, agents, terminal output, agent messaging, or waits.
espennilsen/pi
Delegate tasks to specialized subagents for parallel or sequential execution.
espennilsen/pi
Generate comprehensive status reports for git repositories. Analyzes commits, codebase structure, type errors, branch health, dependency status, GitHub PRs/issues/CI, working tree changes, and optionally integrates with `td` (task management CLI) if a `.todos` folder is present. **TRIGGERS - Use this skill when:** - User asks for the status of a git project or repo - User asks "what's going on in [project]" or "give me an overview of [repo]" - User wants a project summary, progress report, or codebase overview - User asks about recent commits, code structure, or project health - User says "check my project" or "project status" **Output:** A rich, structured text report covering overview, branches, commit history, codebase breakdown, GitHub activity, dependency health, issues/errors, working tree changes, task management, and summary.
espennilsen/pi
Generate a self-contained handoff prompt that lets the user continue work in a fresh coding agent session with empty context. The prompt captures everything a new agent needs: project, goal, progress, decisions, current state, and remaining work. **Triggers — use this skill when:** - User says "handoff", "hand off", "hand-off" - User asks to "generate a continuation prompt" - User says "give me a prompt to continue this" - User asks to "wrap up for a new session" - User says "context transfer", "session transfer" - User asks "how do I pick this up later" - User wants to "pass this to another agent"
espennilsen/pi
Monitor open GitHub PRs across all repos in ~/Dev. Checks PR age, flags stale PRs, and optionally alerts via Aivena/Telegram. Uses `gh pr list` for accurate GitHub data. **TRIGGERS - Use this skill when:** - User asks about open PRs, pull requests, or PR status - User says "check PRs", "any stale PRs?", "PR review status" - Cron job `pr-monitor` fires - User asks "what needs review?" or "what's waiting for merge?" - User wants a PR dashboard or PR age report
espennilsen/pi
Turn one user criticism into a principle-level fix: find the violated principle, every similar occurrence, and fix the shared cause. Use only when the user explicitly invokes expand, $expand, /up:expand, or /skill:expand.
btseytlin/ultrapack
Skills for improving and maintaining skills
rossoctl/rossoctl
Execute this skill optimizes prompts for large language models (llms) to reduce token usage, lower costs, and improve performance. it analyzes the prompt, identifies areas for simplification and redundancy removal, and rewrites the prompt to be more conci... Use when optimizing performance. Trigger with phrases like 'optimize', 'performance', or 'speed up'.
jeremylongshore/tons-of-skills-marketplace
Execute this skill enables AI assistant to create intelligent alerting rules for proactive performance monitoring. it is triggered when the user requests to "create alerts", "define monitoring rules", or "set up alerting". the skill helps define thresholds, rou... Use when generating or creating new content. Trigger with phrases like 'generate', 'create', or 'scaffold'.
jeremylongshore/tons-of-skills-marketplace
Configure and verify provider prompt caching in ORC nodes and composed trees, including shared delegates, model changes, concurrent users, warming and cache misses. Use for reusable prompt prefixes and inference cost; application result caches are a different mechanism.
ObneyAI/orc
Documents manifest locations for an example skill.
NVIDIA/SkillSpector
A friendly greeting skill that says hello to users
NVIDIA/SkillSpector
A helpful assistant agent
NVIDIA/SkillSpector
Orchestrate unattended resolution of open GitHub issues - triage by complexity, dispatch parallel worker agents into isolated worktrees, review their diffs, open PRs, and monitor CI and CodeRabbit until green. Use when asked to work through open issues autonomously or put Claude on auto-pilot.
Jgocunha/dynamic-neural-field-composer
Use when a support or sales bot on a live website must behave: persona/system prompt, grounding so it cannot invent prices or policy, jailbreak and injection defense, the human handoff, launch metrics and kill switch. NOT the agent loop or RAG index under it (that is `building-agents`), NOT a human answering one ticket (that is `customer-support`).
ericrisco/rsc-harness
Plan a huge chunk of work — more than one agent session can hold — as a shared map of decision tickets on your issue tracker, and resolve them one at a time until the way to the destination is clear.
egma-ai/egma
Use when an AI agent's draft response needs a pre-flight check for completion overclaiming, unsupported numeric claims, relayed authority, or unhedged certainty before it ships — or when past chat-log corrections should carry forward into a similar task. Deterministic, offline, zero-LLM gate plus a corrections-memory layer.
hermes-labs-ai/hermeneutic
Survey one surface of tdmcp (artist controls, library/packaging, CLI/DX, AI/LLM, or TouchDesigner depth) for candidate NEW features — inventory what exists, cross-check the roadmap, apply the gap-finding lenses, vet each idea, and emit a structured, novelty- and confidence-labelled candidate list to _workspace/discovery/. Use when a td-surveyor agent is scouting a surface during the feature-discovery harness.
hybridlabor-api/aos