BBuf/AI-Infra-Auto-Driven-SKILLS

---

name: model-pr-history-knowledge

description: Use when an SGLang, vLLM, TensorRT-LLM, or TokenSpeed serving/model optimization task needs prior model-family PR evidence. Query and read the PR-driven history docs under model-pr-optimization-history before choosing source paths, fast paths, kernel/fusion ideas, regression risks, or validation lanes.

---

name: llm-pipeline-analysis

description: Inspect LLM torch profiler traces at forward-pass, layer, and kernel level. Use when you need layer timings, anchor-kernel boundaries, representative kernel flows, or Perfetto time ranges.

---

name: llm-serving-auto-benchmark

description: Framework-independent LLM serving benchmark skill for comparing SGLang, vLLM, TensorRT-LLM, TokenSpeed, or another serving framework. Use when a user wants to find the best deployment command for one model across multiple serving frameworks under the same workload, GPU budget, and latency SLA.

---

name: llm-serving-capacity-planner

description: Parse SGLang/vLLM startup logs to explain GPU memory use and request capacity. Use for KV cache budget, mem-fraction-static comparisons, OOM triage, and max-concurrency estimates.

---

name: llm-torch-profiler-analysis

description: Analyze existing Torch Profiler Chrome traces or capture from SGLang, vLLM, TensorRT-LLM and TokenSpeed servers. Return kernel attribution, overlap opportunities and fusion patterns; use for model dispatch, CUDA Graph gaps, stream contention or exposed kernel tails. Includes existing SGLang Omni traces.

---

name: model-architecture-diagram

description: Return public original model architecture diagrams for user-specified LLM, VLM, MoE, diffusion, OCR, and SGLang/sgl-cookbook model families. Use when the user asks for a model structure chart, architecture diagram, or rendered image link for a specific model such as DeepSeek, GLM, Qwen, Kimi, MiniMax, Step, Hunyuan, or Qwen3-VL.

---

name: model-compute-simulation

description: Build an operator-level compute template for an LLM and estimate FLOPs/MFU for a serving shape. Use when you need tensor shapes, per-op FLOPs, kernel-to-op MFU mapping, or parallelism what-if analysis.

---

name: sglang-model-day0-support

description: Build or audit an evidence-driven SGLang Day-0 support program for a new LLM, VLM, MoE, hybrid-attention, or speculative-decoding model. Use when Codex needs to map a model architecture into SGLang runtime work, design a public support PR DAG, create validation and release gates, sanitize private development evidence, distinguish Day-0 requirements from later fixes or optimizations, or review whether an existing model-support PR is release-ready.

---

name: sglang-explainer-video

description: Turn an SGLang PR (number or URL) or a technical article (Markdown path) into a Chinese explainer video that fully explains every algorithm and code change (length follows the material, no time cap) — natural edge-tts narration, word-synced bilingual 中文/English captions, the unified orange "SGLang PR 讲解 / SGLang 技术解读" cover, SGLang-brand visuals — built on HyperFrames' pr-to-video pipeline and delivered to ~/Desktop. Use when the user sends an sgl-project/sglang PR number or link, or an article file, and asks for a video (做个视频, 讲解视频, PR 视频, 文章转视频).

---

name: sglang-humanize-review

description: Perform SGLang code review in the style of human maintainers by consulting the public PR review episode corpus from project start through 2026-07-27, including inline review threads, top-level PR comments, review submissions, original multilingual text, and multi-round discussions. Use when reviewing SGLang PRs, diffs, patches, or local changes for correctness, tests, performance, GPU/runtime risks, API compatibility, and maintainability.

---

name: sglang-prod-incident-triage

description: Replay-first debug flow for SGLang serving problems. Use when a live or recent server shows health-check failures, latency or throughput regressions, queue growth, timeouts, distributed stalls, crash dumps, wrong outputs after deploys, or PD/EP/HiCache issues, and the job is to turn the problem into a replay plus the right next debug tool.

---

name: torch-profiler-layer-track

description: Add verified layer-number guides and compact GPU display lanes to an existing Torch Profiler Chrome JSON trace. Use for L0/L1/... navigation or traces with hundreds of CUDA Graph stream rows. Changes visualization, not runtime streams.