Plain SKILL.md directories that give a coding agent the operational memory
for real AI-infra work: fair cross-framework benchmarks, kernel-level profiler
reads, operator FLOPs, model support plans, and the upstream PRs that already
solved a similar problem. Kernel campaigns live in the sibling
KDA-Pilot; per-model diffusion runs
live in sglang-diffusion-optimization-flows/.

Releases

v0.1.5
— Release notes (中文), covering model Day-0 support,
profiler layer guides, 118 bilingual model histories, and the October source audit.
Previous release: v0.1.0.

Skills

Model PR History

model-pr-optimization-history/ is one
queryable knowledge base (installed as model-pr-history-knowledge) with
118 bilingual dossiers: SGLang 45, vLLM
44, TensorRT-LLM 15, TokenSpeed 14.
Each lists a model family's implementation files, every PR that changed them,
and per-PR evidence cards. New generated entries are explicitly marked as
source inventories pending manual diff review. Read it before
patching a model path or calling an optimization new.

cd model-pr-optimization-history
python3 scripts/query.py --list
python3 scripts/query.py --framework vllm "qwen3 fused qk norm"

Dossiers are regenerated from upstream git history with
tools/rebuild_model_pr_history_from_git.py; see
update_prompt.md for the full refresh procedure and
docs/upstream-source-contracts.md for
the inspected source revisions.

Install

Claude Code plugin:

/plugin marketplace add BBuf/AI-Infra-Auto-Driven-SKILLS
/plugin install ai-infra-auto-driven-skills@ai-infra-auto-driven-skills

Any skill runtime (Claude Code, Codex, Kimi, ...): link or copy the skill
directories into its skill directory.

git clone https://github.com/BBuf/AI-Infra-Auto-Driven-SKILLS.git
cd AI-Infra-Auto-Driven-SKILLS
SKILL_DIR=~/.claude/skills   # or ${CODEX_HOME:-~/.codex}/skills
mkdir -p "$SKILL_DIR"
for d in skills/*/ skills/model-optimization/*/; do
  [ -f "$d/SKILL.md" ] && ln -sfn "$PWD/${d%/}" "$SKILL_DIR/$(basename "$d")"
done
ln -sfn "$PWD/model-pr-optimization-history" "$SKILL_DIR/model-pr-history-knowledge"

Evidence Rules

• Benchmark rows record model, framework commit, GPUs, workload, rate or
  concurrency, SLA status, both commands, and raw artifacts.
• Profiler reports keep prefill and decode separate and never reuse an older
  trace for a new capture.
• Performance claims are scoped to the exact model, hardware, precision,
  workload, and framework revisions; accuracy is checked on the real path.
• Historical PR evidence keeps its audit date; a source refresh is not a GPU
  rerun.

Related Projects

• KDA-Pilot hosts standalone kernel
  loops, kernel knowledge, and NCU workflows.

Star History

{
  "name": "ai-infra-auto-driven-skills",
  "description": "Agent-ready LLM serving, SGLang model Day-0 support, profiling, capacity, incident triage, architecture, and PR-history skills.",
  "owner": {
    "name": "BBuf",
    "url": "https://github.com/BBuf"
  },
  "plugins": [
    {
      "name": "ai-infra-auto-driven-skills",
      "source": "./",
      "description": "LLM serving benchmarks, SGLang model Day-0 support, profiling, code review, prod incident triage, and PR-history dossiers.",
      "version": "0.8.0",
      "category": "ai-infra",
      "tags": [
        "sglang",
        "day0",
        "vllm",
        "tensorrt-llm",
        "benchmarks",
        "profiler",
        "kernel",
        "moe",
        "llm-serving"
      ]
    }
  ]
}