Saddss
- 38 skills
- 0 followers
- 10 hours ago last updated
- ▌ PDF · saddss bundleUse this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text/tables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting/decrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill.
- ▌ Tdd · saddss bundleTest-driven development with red-green-refactor loop. Use when user wants to build features or fix bugs using TDD, mentions "red-green-refactor", wants integration tests, asks for test-first development, or writes tests for benchmark harnesses, profiling scripts, or inference automation.
- ▌ DOCX · saddss bundleUse this skill whenever the user wants to create, read, edit, or manipulate Word documents (.docx files). Triggers include: any mention of 'Word doc', 'word document', '.docx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads. Also use when extracting or reorganizing content from .docx files, inserting or replacing images in documents, performing find-and-replace in Word files, working with tracked changes or comments, or converting content into a polished Word document. If the user asks for a 'report', 'memo', 'letter', 'template', or similar deliverable as a Word or .docx file, use this skill. Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation.
- ▌ PPTX · saddss bundleUse this skill any time a .pptx file is involved in any way — as input, output, or both. This includes: creating slide decks, pitch decks, or presentations; reading, parsing, or extracting text from any .pptx file (even if the extracted content will be used elsewhere, like in an email or summary); editing, modifying, or updating existing presentations; combining or splitting slide files; working with templates, layouts, speaker notes, or comments. Trigger whenever the user mentions "deck," "slides," "presentation," or references a .pptx filename, regardless of what they plan to do with the content afterward. If a .pptx file needs to be opened, created, or touched, use this skill.
- ▌ Review · saddssTwo-axis branch review — Standards (CONTRIBUTING.md, CONTEXT, ADRs) and Spec (issue/PRD/plan). Runs parallel sub-agents on git diff. Use before PR, when reviewing a branch, or after implementing routing/benchmark/harness changes.
- ▌ Handoff · saddssCompact the current conversation into a handoff document for another agent to pick up. Use when ending a long session, switching agents, or handing off benchmark runs, profiling investigations, or multi-step infra work.
- ▌ Diagnose · saddss bundleDisciplined diagnosis loop for hard bugs and performance regressions. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, describes a performance regression, or debugs vLLM / SGLang / TRT-LLM / lmdeploy failures, benchmark regressions, or flaky inference runs.
- ▌ Prototype · saddss bundleBuild throwaway prototypes to validate a design before production code. Terminal state-machine for routing/cache/overload logic, or UI variants on one route. Use when prototyping, sanity-checking a state model, exploring routing policies, or saying "prototype this" / "let me play with it".
- ▌ To Issues · saddssBreak a plan, spec, or PRD into independently-grabbable GitHub issues using tracer-bullet vertical slices. Use when converting a plan into issues, splitting a benchmark study / routing feature / multi-PR infra work into tickets, or preparing AFK-agent-ready tasks.
- ▌ Search First · saddssResearch-before-coding workflow. Search the repo, framework APIs, libraries, and existing patterns before writing custom code. Use when adding a feature to a mature framework (vLLM, SGLang, TRT-LLM, benchmark harness), before creating a utility/helper, or when the user asks to "add X functionality".
- ▌ Brainstorming · saddss bundleUse before any creative work — adding features, modifying behavior, or extending a mature framework (vLLM, SGLang, TRT-LLM, benchmark harness). Explores intent, constraints, and design; requires user approval before code.
- ▌ Perf Analysis · saddssPerformance analysis coordination workflow. Guides profiling delegation, bottleneck classification (compute / memory / launch / communication / sync), and structured report generation. Use when the user asks to analyze performance, profile a workload, check MFU / SOL, diagnose bottlenecks, understand why a vLLM / SGLang / TRT-LLM serve / lmdeploy run is slower than expected, or interpret an existing .nsys-rep against an SLO. Triggers also include "分析性能瓶颈" / "MFU 多少" / "瓶颈是 compute 还是 memory" / "诊断推理慢"
- ▌ Simplify Code · saddssSimplify and refine recently changed code for clarity, reuse, and efficiency while preserving behavior. Use before opening a PR, after implementing a feature in a mature framework, or when asked to clean up a branch diff.
- ▌ Writing Plans · saddssWrite bite-sized implementation plans from an approved spec. Use after brainstorming or grill-with-docs, before touching code on multi-step features in mature codebases (vLLM, SGLang, inference harness, serving stack patches).
- ▌ Executing Plans · saddssExecute a written implementation plan task-by-task with verification checkpoints. Use after writing-plans when implementing multi-step features in mature frameworks; stop when blocked instead of guessing.
- ▌ Grill With Docs · saddss bundleGrilling session that challenges a plan against the existing domain model, sharpens terminology, and updates documentation (CONTEXT.md, ADRs) inline as decisions crystallise. Use when user wants to stress-test a plan against project language and documented decisions, or before major inference-stack or benchmark architecture changes.
- ▌ Perf Optimization · saddssPerformance optimization coordination playbook. Contains specialist routing table, TileIR two-step pipeline, kernel generation specialist selection, prioritization criteria, and safe modification workflow. Use when applying optimizations, writing kernels, or improving vLLM / SGLang / TRT-LLM / lmdeploy performance. Covers user-specified and autopilot-driven iterative optimization.
- ▌ Parallel Exploring · saddssExplore a large codebase in parallel by launching multiple explore subagents that each investigate a different area simultaneously. Use when onboarding onto a new project, understanding architecture, investigating cross-cutting concerns, or mapping vLLM / SGLang / TRT-LLM / FlexKV code paths.
- ▌ Perf Host Analysis · saddss bundleAnalyze host/CPU overhead in LLM inference (vLLM / SGLang / TRT-LLM serve / lmdeploy / TensorRT-LLM) from nsys traces. Phase 1 (Detection): determine whether host overhead is the bottleneck via a binary YES/NO verdict with metric evidence (GPU idle ratio, host prep exposed ratio, per-phase breakdown). Phase 2 (Root Cause): isolate forward step iterations via allreduce kernel patterns, compare NVTX-instrumented host operations across versions, and pinpoint scheduling / request-management regressions. Usable standalone or as a sub-step of perf-analysis. Triggers: host overhead, inter-step gap, scheduling overhead, forward step isolation, nsys iteration analysis, NVTX breakdown, request management overhead, inference loop overhead, between-step gap, GPU idle, host bottleneck detection, host prep exposed, "GPU 利用率上不去", "Python 调度太慢", "vLLM scheduler overhead".
- ▌ Codebase Onboarding · saddssLaunch multiple explore subagents in parallel to investigate architecture, data models, APIs, and deployment; synthesize an onboarding document. Use when first entering an inference framework repo, benchmark harness, or unfamiliar serving stack.
- ▌ Kernel Cute Writing · saddss bundleWrite GPU kernels using NVIDIA CuTe DSL (CUTLASS 4.x Python API) — NOT for Triton, CUDA C++, or conceptual-only questions. Covers element-wise kernels, GEMM patterns, reductions, memory hierarchy (global/shared/register/TMA), MMA tensor core ops, software pipelining, and framework integration. Use when implementing CUTLASS/CuTe kernels for inference GEMM, MoE, or attention in TRT-LLM / custom serving stacks.
- ▌ Perf Nsight Systems · saddss bundleNsight Systems (nsys) CLI for system-level timeline profiling. Use when the user wants to run nsys profile, analyze .nsys-rep reports, use nsys stats/analyze/recipe commands, diagnose GPU idle time from timeline traces, profile distributed training with NCCL overlap analysis, or profile an LLM inference server (vLLM / SGLang / TRT-LLM serve / lmdeploy) and its client (e.g. online_replay.py) end-to-end. Triggers also include "用 nsys 抓 profile" / "看 GPU 空闲" / "排查 e2e 延迟" / "分析 NCCL overlap" / "trace 推理服务". NOT for kernel-level metrics like SOL%, occupancy, or roofline (use perf-nsight-compute-analysis for ncu). NOT for writing or generating kernels. NOT for applying optimizations like CUDA Graphs.
- ▌ Kernel Triton Writing · saddss bundleOpenAI Triton (@triton.jit) kernel development for inference/training ops. NEVER use for CUDA C++, TileIR, or profiling tools (ncu, nsys). Covers fused elementwise, reductions (softmax, LayerNorm, RMSNorm), tiled GEMM with triton.autotune, and flash attention. Workflow: design, write, verify. Use when implementing custom attention, MoE, or KV-cache kernels in vLLM / SGLang stacks. Triggers include "write triton kernel", "Triton GEMM".
- ▌ LLM Pipeline Analysis · saddss bundleLayer-level torch profiler trace analysis: forward-pass boundaries, anchor kernels, per-layer timing clusters, Perfetto navigation. Use when llm-torch-profiler-analysis three-table summary is too coarse and you need which layers or prefill/decode forward passes dominate.
- ▌ Perf Host Optimization · saddss bundleProfiles and optimizes LLM inference host / CPU overhead (vLLM / SGLang / TRT-LLM serve / lmdeploy / TensorRT-LLM PyExecutor) using line_profiler (with nsys support planned). Runs iterative profile-analyze-optimize-validate rounds. Use when GPU utilization is low, when perf-host-analysis has concluded "host-bound", or when the user asks to reduce host overhead / Python scheduling overhead / sampler latency / inference loop overhead. Triggers also include: "削 host overhead", "降 launch overhead", "优化 scheduler", "GPU 利用率低怎么提".
- ▌ Perf Workload Profiling · saddss bundleCode instrumentation for timing workloads. Two scenarios: (1) Training loop — inject manual timing to report per-iteration latency, throughput (samples/sec), and data load time. (2) Standalone kernel/op — write CUDA event timing code with warmup, per-iteration statistics, and anti-pattern avoidance. Also covers NVTX annotation for labeling profiler timelines. NOT for: running or analyzing profiler tools (nsys, ncu, Nsight Systems, Nsight Compute), writing kernels (Triton, CuTe, CUDA), applying optimizations (CUDA Graphs, gradient checkpointing, fusion), or interpreting roofline/SOL% metrics. Triggers: "measure throughput", "benchmark this function", "time my training loop", "samples per second", "NVTX annotate", "instrument my dataloader", "data load time", "kernel timing", "how do I time", inference step timing, vLLM forward latency instrumentation.
- ▌
- ▌ Academic Figure Workflow · saddss bundleCreate, revise, and validate publication-style academic figures for papers, theses, and technical reports, especially editable SVG mechanism/architecture diagrams and their DOCX/WPS insertion. Use for 论文作图, 学术架构图, 机理图, SVG 绘图, 图片重排, 文字/箭头对齐, 图注, 600 dpi 预览, 图片插入 Word/WPS, or when repeated visual defects must be eliminated with deterministic QA. Do not use for photorealistic illustration or decorative presentation slides.
- ▌ Model Compute Simulation · saddss bundleBuild operator-level compute templates from model config; estimate FLOPs, MFU, tensor shapes, and TP/EP what-if for serving shapes. Use before benchmarking to sanity-check compute bounds or when writing REPORT.md theoretical sections.
- ▌ Model Perf Binary Search · saddss bundleFind the maximum sustainable QPS of an LLM inference service that meets a p50 e2e latency SLO using online_replay.py and a binary search. Use for maximum-QPS benchmarks, SLO-based performance tests, and optional serving configuration or feature tuning on local OpenAI-compatible servers.
- ▌ Setup Matt Pocock Skills · saddss bundleScaffold docs/agents/ in a target repo — issue tracker (GitHub/GitLab/local), triage labels, domain doc layout. Run once per repo before to-issues, review, or other skills that read docs/agents/issue-tracker.md.
- ▌ LLM Serving Auto Benchmark · saddss bundleFramework-independent LLM serving benchmark for SGLang, vLLM, TensorRT-LLM. Config-driven search_space over launch flags under shared workload, GPU budget, and latency SLA. Use for cross-framework deployment comparison, cookbook sweeps, or finding best serve command before QPS binary search. Pairs with model-perf-binary-search for SLO-max-QPS tuning after a winner is chosen.
- ▌ LLM Torch Profiler Analysis · saddss bundleUnified torch.profiler triage for sglang, vllm, TensorRT-LLM: kernel table, overlap-opportunity table, fuse-pattern table with prefill/decode separation. Use when analyzing trace.json(.gz), live server profiling, or fusion/kernel opportunities after perf-nsight-systems points to compute. Complements llm-pipeline-analysis for layer-level drill-down.
- ▌ High Performance Code Review · saddssReview pull requests and code changes for high-performance Python, PyTorch, CUDA, and distributed systems using a correctness-first P0-P4 rubric. Use when the user asks for a strict production-grade review, especially for inference, GPU, concurrency, or performance-sensitive code.
- ▌ LLM Serving Capacity Planner · saddss bundleParse SGLang/vLLM startup logs for GPU memory decomposition (weights, KV pool, CUDA graph, overhead) and max-concurrency estimates. Use for KV cache budget, mem-fraction-static tuning, OOM triage, or explaining why concurrency is capped.
- ▌ Perf Nsight Compute Analysis · saddss bundleAnalyze ncu (NVIDIA Nsight Compute) profiling output: SOL% bottleneck classification, roofline analysis, occupancy diagnosis, memory hierarchy analysis, warp stall analysis, metric interpretation, and programmatic .ncu-rep report analysis. Use when optimizing CUDA kernels in vLLM / SGLang / TRT-LLM / custom inference ops, or interpreting .ncu-rep for attention / GEMM / KV-cache kernels. NOT for kernel writing, nsys timeline work, or host-side profiling. Triggers also include "SOL 多少" / "ncu 分析".
- ▌ Improve Codebase Architecture · saddss bundleFind deepening opportunities in a codebase, informed by the domain language in CONTEXT.md and the decisions in docs/adr/. Use when the user wants to improve architecture, find refactoring opportunities, consolidate tightly-coupled modules, make a codebase more testable and AI-navigable, or untangle inference engine / benchmark harness code.
- ▌ Verification Before Completion · saddssBefore claiming work is complete, fixed, or passing — run verification commands and cite evidence. Use before commit/PR, after implementing features or benchmark fixes, when saying tests pass or bug is fixed.