vllm
Route vLLM tasks across offline inference, OpenAI-compatible serving, structured/tool/reasoning, multimodal/LoRA/pooling, and deployment/performance workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- vllm
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Route vLLM tasks across offline inference, OpenAI-compatible serving, structured/tool/reasoning, multimodal/LoRA/pooling, and deployment/performance workflows.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/repo-provenance.md2.1 KB
- 参考文档references/repo-routing-metadata.json432 B
- 参考文档references/troubleshooting.md2.9 KB
- 脚本scripts/vllm_skill_doctor.py3.2 KB
03 SKILL.md 正文
vLLM
Use this skill when the user asks how to install, use, serve, configure, troubleshoot, or optimize vLLM, the high-throughput inference and serving engine for LLMs. This root skill is a router; open the focused sub-skill before writing substantial commands or code.
First Checks
- Confirm whether the user wants in-process Python inference or an OpenAI-compatible HTTP server.
- Ask for the model identifier/path, hardware target, and whether model downloads, remote code, local media, or credentials are allowed before running live model commands.
- Treat GPU execution, model downloads, multi-node launches, and benchmark runs as environment-gated; prefer help/static checks until the user confirms resources.
- Use
references/repo-provenance.mdto check whether this skill is aligned with the current vLLM checkout. - Use
references/troubleshooting.mdfor cross-cutting install/import/backend triage before routing to workflow-specific troubleshooting.
Route by Workflow
| User asks for | Open |
|---|---|
Python LLM, SamplingParams, generate, chat, encode, embed, classify, score, output extraction, or offline smoke scripts |
sub-skills/offline-inference/SKILL.md |
vllm serve, OpenAI clients, /v1 endpoints, chat, complete, run-batch, model discovery, server/client errors |
sub-skills/openai-serving/SKILL.md |
| Structured outputs, JSON schema, regex/grammar constraints, tool calling, reasoning parsers, chat templates, streaming tool-call deltas | sub-skills/structured-tool-reasoning/SKILL.md |
| Multimodal image/audio/video payloads, media allowlists, LoRA/adapters, embeddings, rerank, score, pooling outputs | sub-skills/modalities-adapters-pooling/SKILL.md |
| Memory planning, parallelism, quantization, KV cache/offload, disaggregated serving, Ray/torchrun, metrics, profiling, benchmarks | sub-skills/deployment-performance/SKILL.md |
Common Starting Points
- Install/import sanity: vLLM supports Python 3.10 through 3.14 in package metadata; runtime builds are backend-specific. For source installs, follow the project’s current documented environment workflow instead of mixing package managers.
- Minimal import check:
python -c "import vllm; print(vllm.__version__)"verifies importability, but it does not prove a model/backend can run. - CLI discovery:
vllm --help,vllm serve --help, andvllm serve --help=allexpose current command groups and flags. - Server default:
vllm serve [model_tag] [options]launches an OpenAI-compatible server; if omitted, current CLI help documents a default small Qwen model, but production commands should pass an explicit model. - Offline default:
LLM(model=...)plusSamplingParams(max_tokens=...)is the main in-process path; chat models usually needLLM.chator an explicit chat template rather than rawgenerateprompts.
Bundled Root Helpers
scripts/vllm_skill_doctor.py: checks Python import, distribution metadata, CLI availability, and selected backend facts without downloading models or starting a server.
Sub-skills also include focused helpers for offline smoke planning, serve command construction, OpenAI client smoke checks, structured request validation, multimodal payload validation, environment summaries, and memory command planning.
Safety Defaults
- Do not start a server, download a model, enable
trust_remote_code, read arbitrary local media, execute tool calls, or run distributed/benchmark commands unless the user explicitly approves the required resources and side effects. - For local media in OpenAI-compatible multimodal requests, require a narrow
--allowed-local-media-path; for remote media, use explicit--allowed-media-domainentries. - For API keys, tokens, Hugging Face credentials, and server auth, describe required environment variables or headers without echoing secret values.
- For performance problems, collect environment and configuration facts first, then propose bounded experiments; avoid broad benchmark suites unless requested.
Verification Scope
This skill was generated from vLLM source, docs, examples, tests, package metadata, and a CPU/precompiled package-inspection environment. Live GPU model serving, model downloads, multi-node runs, and native benchmark execution remain user-environment gated and should be verified with the user’s actual hardware and model.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).