The missing manual for running model X on hardware Y

What it does vLLM Recipes hosts community-maintained cookbooks that translate the generic inference engine into concrete configurations for dozens of models, from DeepSeek and Qwen to Llama and Gemma. Each recipe answers the specific question of how to run a given model on a given hardware setup. The structured YAML files render into a browsable site at recipes.vllm.ai.

The interesting bit Instead of scattered README tips, the project is migrating toward a validated YAML pipeline with a formal schema and a VRAM formula. That turns tribal knowledge about GPU memory into reproducible, checkable recipes.

Key highlights

• Covers a broad model zoo including DeepSeek, Qwen, GLM, MiniMax, and NVIDIA Nemotron.
• Recipes live as structured YAML under models/<hf_org>/<hf_repo>.yaml with a defined schema.
• Includes validation scripts and a VRAM formula to sanity-check configurations before you deploy.
• Renders to a dedicated site rather than forcing users to hunt through GitHub directories.
• Legacy Markdown guides remain as reference during the ongoing YAML migration.

Caveats

• The repository is actively migrating formats; top-level Markdown folders are the old MkDocs site and may drift out of sync with the new YAML source.
• Coverage depends on community contributions, so newer or niche models may not yet have recipes.

Verdict Keep this bookmarked if you run vLLM across a changing model zoo and need quick memory math. If you only ever serve one well-known model, you probably already know the magic flags.