AI BriefingvLLM Recipes Revamp - One Click for Model+Hardware Combo Configs
·2026.04.23 13:42
Key point
vLLM Recipes has been revamped to support hardware-specific serving configs and command generation.
Details
recipes.vllm.ai has undergone a major revamp, making it possible to directly find and generate how to run model X on hardware Y right on the web.
The key changes are as follows.
- HuggingFace mirror URL: By replacing
huggingface.cowithrecipes.vllm.ai, you can go straight to that model's recipe page. Example:recipes.vllm.ai/Qwen/Qwen3.6-35B-A3B - Interactive command builder: Selecting hardware, variant, and parallelization strategy generates a ready-to-run
vllm servecommand. - Pluggable hardware support: Switch between NVIDIA Hopper/Blackwell and AMD MI300X/MI355X with one click, with related flags and environment variables automatically applied.
- JSON API provided: All recipes are served at the
/.jsonpath, allowing agents or automation tools to consume them directly.
The repository also includes agent skills, enabling contributions that flow from running benchmarks to generating recipes to submitting PRs.
The repository is at github.com/vllm-project/recipes.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
vLLM v0.28.0 Release: Kimi-K3 and DeepSeek V4 Optimizations, Significant Inference Performance Improvements