vLLM Recipes

This repo intends to host community maintained common recipes to run vLLM answering the question: How do I run model X on hardware Y for task Z?

Guides

Arcee AI

• Trinity-Large-Thinking

DeepSeek

• DeepSeek-OCR
• DeepSeek-V3.1
• DeepSeek-V3, DeepSeek-R1
• DeepSeek-V3.2-Exp
• DeepSeek-V3.2

Ernie

• Ernie4.5
• Ernie4.5-VL

GLM

• GLM-5.1, GLM-5
• Glyph
• GLM-4.7-Flash, GLM-4.7, GLM-4.6, GLM-4.5, GLM-4.5-Air
• GLM-4.5V, GLM-4.6V

Google

• Gemma 4
• TranslateGemma

inclusionAI

• Ring-1T-FP8

InternVL

• InternVL3.5

InternLM

• Intern-S1

Jina AI

• Jina-reranker-m0

Llama

• Llama4-Scout
• Llama3.3-70B
• Llama3.1

Microsoft

• Phi-4

MiniMax

• MiniMax-M2.7, MiniMax-M2.5, MiniMax-M2.1, MiniMax-M2

Xiaomi MiMo

• MiMo-V2-Flash

Mistral AI

• Ministral-3-Instruct
• Ministral-3-Reasoning
• Mistral-Large-3-Instruct

Moonshotai

• Kimi-K2.5
• Kimi-K2
• Kimi-K2-Think
• Kimi-Linear

NVIDIA

• Nemotron-3-Nano-30B-A3B
• Nemotron-Nano-12B-v2-VL

OpenAI

• gpt-oss

PaddlePaddle

• PaddleOCR-VL

Qwen

• Qwen3
• Qwen3-ASR
• Qwen3-VL
• Qwen3-Next
• Qwen3.5
• Qwen3-Coder-480B-A35B
• Qwen2.5-VL
• Qwen3Guard-Gen

Seed

• Seed-OSS-36B

StepFun

• Step-3.5-Flash

Tencent-Hunyuan

• HunyuanOCR

How recipes are maintained

Each recipe is a structured YAML file at models/<hf_org>/<hf_repo>.yaml, kept in the repository of the project that knows the model best. This repository is the renderer: it pulls those files in, merges them, and builds recipes.vllm.ai and its JSON API.

flowchart LR
  vllm["vllm-project/vllm<br/>recipes/models/"] --> sync
  omni["vllm-project/vllm-omni<br/>recipes/models/"] --> sync
  ascend["vllm-ascend<br/>overlays/"] -.-> sync
  self["this repo<br/>taxonomy · strategies · kv_store"] --> assemble
  sync["sync-sources.mjs<br/>pinned by sources.lock.json"] --> assemble["assemble-content.mjs<br/>→ .content/&lt;channel&gt;/"]
  assemble --> build["build-recipes-api.mjs + next build"] --> site["recipes.vllm.ai<br/>+ JSON API"]

Content	Where it lives	Status
Model recipes served by vLLM	vllm-project/vllm → recipes/models/	Owned by vLLM; a merge there reaches recipes.vllm.ai in about 30 minutes. Copies stay in this repo's models/ until they are synced to vLLM and removed; they aren't rendered
Omni recipes (image, video, audio generation)	vllm-project/vllm-omni → recipes/models/	Owned by vLLM-Omni, reaching recipes.vllm.ai the same way. Copies stay in this repo's models/ until they are removed; they aren't rendered
Ascend NPU content	Inline in the vLLM recipes	Moving to overlays/ in vllm-ascend
Schema, hardware profiles, serving strategies, KV-store deployments, platforms	This repo (taxonomy.yaml, strategies/, kv_store/, platforms.yaml)	Owned here
Command synthesis, site, JSON API	This repo (src/, scripts/)	Owned here

How the pieces fit:

• sources.yaml lists every source repository for a channel (today only latest). For each source it gives the repo, ref, and directory, a role, and the scope this repo grants it:
• core / owner sources ship whole recipes. Each Hugging Face ID may come from only one source.
• overlay sources add content to another source's recipe, for example Ascend profiles, variants, and overrides from vllm-ascend. An overlay can only add keys, and only for the hardware brands its scope grants, so it can never change what the NVIDIA or AMD commands render.
• sources.lock.json pins every remote source to a commit, so each build is reproducible.
• Updates are automatic. bump-sources.yml runs every 20 minutes: 1. Checks whether a source's recipe directory changed. 2. Validates the new content with the full JSON API build. 3. Commits the lock, which redeploys the site.

If validation fails, the lock stays on the last good commit, so the site keeps serving the previous content, and a sources-bump-failed issue is opened. There is one issue per failure, not one per run, and it closes itself on the next successful run. - Provenance. Every recipe page links "Edit recipe" to the repository that owns the file. Every recipe JSON carries a source field (repo, commit, path) for tracing a command back to its YAML.

Contributing

Edit or add a recipe in the repository that owns it (see the table above), following CONTRIBUTING.md for the full schema, VRAM formula, and validation steps. Changes to the schema, hardware profiles, strategies, or the site go to this repository.

To preview a recipe change from a local checkout of a source repository, create a gitignored sources.local.yaml in this repo:

overrides:
  vllm: ~/vllm/recipes
  vllm-omni: ~/vllm-omni/recipes

Quick loop:

pnpm install
pnpm dev                              # http://localhost:3000
node scripts/sync-sources.mjs         # fetch the source commits pinned in sources.lock.json
node scripts/build-recipes-api.mjs    # assembles sources, validates every YAML, rebuilds the JSON API
node scripts/validate-source.mjs --source vllm --dir ~/vllm/recipes   # validate a source checkout (what its CI runs)
pnpm content:watch                    # re-assemble when files are added to or removed from a local source

Legacy MkDocs guides

The top-level Markdown directories (DeepSeek/, Qwen/, etc.) are the historical MkDocs site, kept as a reference during the YAML migration. To preview them:

uv venv
source .venv/bin/activate
uv pip install -r requirements.txt
uv run mkdocs serve --dev-addr 127.0.0.1:8001

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.