What problem does it solve?
Adding a new model to the SGLang Cookbook documentation requires coordinated edits across many files — a per-model JSX config, benchmarks data, an MDX page, docs.json navigation, NEW-tag hygiene, and homepage vendor cards — and doing this by hand is error-prone and inconsistent.
Core Features & Use Cases
- Template instantiation: Copies model-agnostic templates (config.jsx, benchmarks.jsx, page.mdx) into per-model files and replaces all
__TOKEN__ placeholders with verified model data.
- Verified deployment matrix: Builds the 5-dimension
cells[] matrix (hardware × variant × quantization × strategy × nodes) from tested sglang serve recipes, with TP sizing sanity checks and platform-specific flags for NVIDIA and AMD GPUs.
- Site wiring and validation: Updates docs.json navigation, enforces one NEW tag per vendor directory, creates homepage vendor cards and logos, then validates with
mint validate, mint broken-links, and check_cookbook_configs.mjs.
- Use Case: When SGLang adds day-0 support for a new model like DeepSeek-V4, use this Skill to generate the complete cookbook page with deployment matrix, playground config, benchmarks, and prose, then open a PR.
Quick Start
Ask the assistant to add a new model to the SGLang Cookbook by providing the model's HuggingFace repo URL and its tested hardware and launch recipes.
