compitabi

https://peterableda.github.io/compitabi/

Answers the question: "In which vLLM version is this HuggingFace model supported?"

Given a model ID like meta-llama/Llama-3.2-11B-Vision-Instruct, it fetches the model's architecture from the HuggingFace API and looks it up against a local cache of vLLM release history.

How it works

1. HuggingFace models declare their architecture in config.architectures (e.g. MllamaForConditionalGeneration)
2. vLLM registers supported architectures as dictionary keys in its model registry files
3. build_cache.py walks every vLLM GitHub release, extracts those architecture names, and records the first version each one appeared in
4. check.py fetches a model's architecture from HuggingFace and looks it up in the cache

Setup

uv sync

Usage

1. Build the cache

Fetches all vLLM releases from GitHub (~100 releases as of 2026). A GITHUB_TOKEN is strongly recommended to avoid the 60 req/hour unauthenticated rate limit.

GITHUB_TOKEN=your_token uv run python build_cache.py

The cache is saved to vllm_cache.json and is incremental — re-running only fetches releases not already cached.

2. Check a model

uv run python check.py meta-llama/Llama-3.2-11B-Vision-Instruct

Example output:

Model:         meta-llama/Llama-3.2-11B-Vision-Instruct
Architectures: MllamaForConditionalGeneration
Cache built:   2026-04-17T10:00:00+00:00

  MllamaForConditionalGeneration  ->  first supported in vLLM v0.10.0

If an architecture is not found, the model is either unsupported or support was added after the cache was last built — re-run build_cache.py to refresh.